tiararodney/sekft
2
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge branch 'develop'

Browse source
This commit is contained in:
Tiara Rodney 2026-06-18 00:59:08 +02:00
commit 705b4a028b
Signed by: tiara
GPG key ID: 5CD8EC1D46106723
18 changed files with 3276 additions and 14 deletions
Download patch file Download diff file Expand all files Collapse all files

41
CHANGELOG.md Normal file
View file

@ -0,0 +1,41 @@
# Changelog
All notable changes to sekft, the shell-operator SFT trainer behind the
[posix-sdc](https://huggingface.co/datasets/tiararodney/posix-sdc) experiment,
are documented in this file.
The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
and the project follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
## [1.0.0] - 2026-06-18
First release: the training and evaluation pipeline that turns posix-sdc
trajectories into a fine-tuned shell operator.
### Added
- `sekft-train`: LoRA / QLoRA supervised fine-tuning of a base model on
shell-operation trajectories, with an **assistant-only loss mask** derived by
token-prefix differencing — the commands and the terminal `exit` / `panic`
token are trained; the environment turns (orientation, prompts, command
output) are masked to `-100`. The render uses the tokenizer's own
`apply_chat_template`, so training matches what the serving harness sends
(train = serve), with `normalize_for_template` canonicalising trajectories for
instruct templates that have no system role and require strict user/assistant
alternation.
- Three sources of training data: a directory of raw rollout `.json`
(keep-filtered), a curated `.jsonl` corpus, or the published posix-sdc corpus
over the Hugging Face Hub (`--hub`).
- `--inspect` for mask and token statistics without training, and structured
stderr logging across every phase (`-v` / `-q`): per-trajectory and progress
lines while the corpus is tokenized, dataset accounting that warns on dropped
(over-length / empty-mask) trajectories, and the per-step training curve.
- `sekft-eval`: behavioural evaluation that drops the tuned model into held-out
scenarios with no scaffold and scores whether it operates and terminates.
- `sekft-resident`: a resident-base harness that loads the base model once and
fits several adapters without reloading, for paired / STaR-style runs.
- Packaging: the `tiararodney.sekft` namespace package with `sekft-train`,
`sekft-eval`, and `sekft-resident` console scripts; a typed (`py.typed`),
mypy-strict codebase; an optional `[gpu]` extra (torch / transformers / peft);
and a dependency on `posix-sdc[hub]`. Released under GPL-2.0.
[1.0.0]: https://git.code.tiararodney.com/tiara/sekft/releases/tag/v1.0.0

14
Dockerfile
View file

@ -1,14 +0,0 @@
# Minimal dash-in-a-box for sekft trajectory generation.
# docker build -t sekft-dash .
#
# dash as the operated shell (strict POSIX, no bashisms), busybox applets for
# the coreutils. busybox is intentionally close to minimal POSIX so trajectories
# transfer toward sek rather than encoding GNU-isms. Add `coreutils findutils
# grep sed` here if you want GNU semantics instead.
FROM alpine:3.19
RUN apk add --no-cache dash \
&& ln -sf /usr/bin/dash /bin/dash \
&& ln -sf /usr/bin/dash /bin/sh
# /work is the default arena; provider files land at their absolute paths.
RUN mkdir -p /work
WORKDIR /work

338
LICENSE Normal file
View file

@ -0,0 +1,338 @@
GNU GENERAL PUBLIC LICENSE
Version 2, June 1991
Copyright (C) 1989, 1991 Free Software Foundation, Inc.,
<https://fsf.org/>
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
Preamble
The licenses for most software are designed to take away your
freedom to share and change it. By contrast, the GNU General Public
License is intended to guarantee your freedom to share and change free
software--to make sure the software is free for all its users. This
General Public License applies to most of the Free Software
Foundation's software and to any other program whose authors commit to
using it. (Some other Free Software Foundation software is covered by
the GNU Lesser General Public License instead.) You can apply it to
your programs, too.
When we speak of free software, we are referring to freedom, not
price. Our General Public Licenses are designed to make sure that you
have the freedom to distribute copies of free software (and charge for
this service if you wish), that you receive source code or can get it
if you want it, that you can change the software or use pieces of it
in new free programs; and that you know you can do these things.
To protect your rights, we need to make restrictions that forbid
anyone to deny you these rights or to ask you to surrender the rights.
These restrictions translate to certain responsibilities for you if you
distribute copies of the software, or if you modify it.
For example, if you distribute copies of such a program, whether
gratis or for a fee, you must give the recipients all the rights that
you have. You must make sure that they, too, receive or can get the
source code. And you must show them these terms so they know their
rights.
We protect your rights with two steps: (1) copyright the software, and
(2) offer you this license which gives you legal permission to copy,
distribute and/or modify the software.
Also, for each author's protection and ours, we want to make certain
that everyone understands that there is no warranty for this free
software. If the software is modified by someone else and passed on, we
want its recipients to know that what they have is not the original, so
that any problems introduced by others will not reflect on the original
authors' reputations.
Finally, any free program is threatened constantly by software
patents. We wish to avoid the danger that redistributors of a free
program will individually obtain patent licenses, in effect making the
program proprietary. To prevent this, we have made it clear that any
patent must be licensed for everyone's free use or not licensed at all.
The precise terms and conditions for copying, distribution and
modification follow.
GNU GENERAL PUBLIC LICENSE
TERMS AND CONDITIONS FOR COPYING, DISTRIBUTION AND MODIFICATION
0. This License applies to any program or other work which contains
a notice placed by the copyright holder saying it may be distributed
under the terms of this General Public License. The "Program", below,
refers to any such program or work, and a "work based on the Program"
means either the Program or any derivative work under copyright law:
that is to say, a work containing the Program or a portion of it,
either verbatim or with modifications and/or translated into another
language. (Hereinafter, translation is included without limitation in
the term "modification".) Each licensee is addressed as "you".
Activities other than copying, distribution and modification are not
covered by this License; they are outside its scope. The act of
running the Program is not restricted, and the output from the Program
is covered only if its contents constitute a work based on the
Program (independent of having been made by running the Program).
Whether that is true depends on what the Program does.
1. You may copy and distribute verbatim copies of the Program's
source code as you receive it, in any medium, provided that you
conspicuously and appropriately publish on each copy an appropriate
copyright notice and disclaimer of warranty; keep intact all the
notices that refer to this License and to the absence of any warranty;
and give any other recipients of the Program a copy of this License
along with the Program.
You may charge a fee for the physical act of transferring a copy, and
you may at your option offer warranty protection in exchange for a fee.
2. You may modify your copy or copies of the Program or any portion
of it, thus forming a work based on the Program, and copy and
distribute such modifications or work under the terms of Section 1
above, provided that you also meet all of these conditions:
a) You must cause the modified files to carry prominent notices
stating that you changed the files and the date of any change.
b) You must cause any work that you distribute or publish, that in
whole or in part contains or is derived from the Program or any
part thereof, to be licensed as a whole at no charge to all third
parties under the terms of this License.
c) If the modified program normally reads commands interactively
when run, you must cause it, when started running for such
interactive use in the most ordinary way, to print or display an
announcement including an appropriate copyright notice and a
notice that there is no warranty (or else, saying that you provide
a warranty) and that users may redistribute the program under
these conditions, and telling the user how to view a copy of this
License. (Exception: if the Program itself is interactive but
does not normally print such an announcement, your work based on
the Program is not required to print an announcement.)
These requirements apply to the modified work as a whole. If
identifiable sections of that work are not derived from the Program,
and can be reasonably considered independent and separate works in
themselves, then this License, and its terms, do not apply to those
sections when you distribute them as separate works. But when you
distribute the same sections as part of a whole which is a work based
on the Program, the distribution of the whole must be on the terms of
this License, whose permissions for other licensees extend to the
entire whole, and thus to each and every part regardless of who wrote it.
Thus, it is not the intent of this section to claim rights or contest
your rights to work written entirely by you; rather, the intent is to
exercise the right to control the distribution of derivative or
collective works based on the Program.
In addition, mere aggregation of another work not based on the Program
with the Program (or with a work based on the Program) on a volume of
a storage or distribution medium does not bring the other work under
the scope of this License.
3. You may copy and distribute the Program (or a work based on it,
under Section 2) in object code or executable form under the terms of
Sections 1 and 2 above provided that you also do one of the following:
a) Accompany it with the complete corresponding machine-readable
source code, which must be distributed under the terms of Sections
1 and 2 above on a medium customarily used for software interchange; or,
b) Accompany it with a written offer, valid for at least three
years, to give any third party, for a charge no more than your
cost of physically performing source distribution, a complete
machine-readable copy of the corresponding source code, to be
distributed under the terms of Sections 1 and 2 above on a medium
customarily used for software interchange; or,
c) Accompany it with the information you received as to the offer
to distribute corresponding source code. (This alternative is
allowed only for noncommercial distribution and only if you
received the program in object code or executable form with such
an offer, in accord with Subsection b above.)
The source code for a work means the preferred form of the work for
making modifications to it. For an executable work, complete source
code means all the source code for all modules it contains, plus any
associated interface definition files, plus the scripts used to
control compilation and installation of the executable. However, as a
special exception, the source code distributed need not include
anything that is normally distributed (in either source or binary
form) with the major components (compiler, kernel, and so on) of the
operating system on which the executable runs, unless that component
itself accompanies the executable.
If distribution of executable or object code is made by offering
access to copy from a designated place, then offering equivalent
access to copy the source code from the same place counts as
distribution of the source code, even though third parties are not
compelled to copy the source along with the object code.
4. You may not copy, modify, sublicense, or distribute the Program
except as expressly provided under this License. Any attempt
otherwise to copy, modify, sublicense or distribute the Program is
void, and will automatically terminate your rights under this License.
However, parties who have received copies, or rights, from you under
this License will not have their licenses terminated so long as such
parties remain in full compliance.
5. You are not required to accept this License, since you have not
signed it. However, nothing else grants you permission to modify or
distribute the Program or its derivative works. These actions are
prohibited by law if you do not accept this License. Therefore, by
modifying or distributing the Program (or any work based on the
Program), you indicate your acceptance of this License to do so, and
all its terms and conditions for copying, distributing or modifying
the Program or works based on it.
6. Each time you redistribute the Program (or any work based on the
Program), the recipient automatically receives a license from the
original licensor to copy, distribute or modify the Program subject to
these terms and conditions. You may not impose any further
restrictions on the recipients' exercise of the rights granted herein.
You are not responsible for enforcing compliance by third parties to
this License.
7. If, as a consequence of a court judgment or allegation of patent
infringement or for any other reason (not limited to patent issues),
conditions are imposed on you (whether by court order, agreement or
otherwise) that contradict the conditions of this License, they do not
excuse you from the conditions of this License. If you cannot
distribute so as to satisfy simultaneously your obligations under this
License and any other pertinent obligations, then as a consequence you
may not distribute the Program at all. For example, if a patent
license would not permit royalty-free redistribution of the Program by
all those who receive copies directly or indirectly through you, then
the only way you could satisfy both it and this License would be to
refrain entirely from distribution of the Program.
If any portion of this section is held invalid or unenforceable under
any particular circumstance, the balance of the section is intended to
apply and the section as a whole is intended to apply in other
circumstances.
It is not the purpose of this section to induce you to infringe any
patents or other property right claims or to contest validity of any
such claims; this section has the sole purpose of protecting the
integrity of the free software distribution system, which is
implemented by public license practices. Many people have made
generous contributions to the wide range of software distributed
through that system in reliance on consistent application of that
system; it is up to the author/donor to decide if he or she is willing
to distribute software through any other system and a licensee cannot
impose that choice.
This section is intended to make thoroughly clear what is believed to
be a consequence of the rest of this License.
8. If the distribution and/or use of the Program is restricted in
certain countries either by patents or by copyrighted interfaces, the
original copyright holder who places the Program under this License
may add an explicit geographical distribution limitation excluding
those countries, so that distribution is permitted only in or among
countries not thus excluded. In such case, this License incorporates
the limitation as if written in the body of this License.
9. The Free Software Foundation may publish revised and/or new versions
of the General Public License from time to time. Such new versions will
be similar in spirit to the present version, but may differ in detail to
address new problems or concerns.
Each version is given a distinguishing version number. If the Program
specifies a version number of this License which applies to it and "any
later version", you have the option of following the terms and conditions
either of that version or of any later version published by the Free
Software Foundation. If the Program does not specify a version number of
this License, you may choose any version ever published by the Free Software
Foundation.
10. If you wish to incorporate parts of the Program into other free
programs whose distribution conditions are different, write to the author
to ask for permission. For software which is copyrighted by the Free
Software Foundation, write to the Free Software Foundation; we sometimes
make exceptions for this. Our decision will be guided by the two goals
of preserving the free status of all derivatives of our free software and
of promoting the sharing and reuse of software generally.
NO WARRANTY
11. BECAUSE THE PROGRAM IS LICENSED FREE OF CHARGE, THERE IS NO WARRANTY
FOR THE PROGRAM, TO THE EXTENT PERMITTED BY APPLICABLE LAW. EXCEPT WHEN
OTHERWISE STATED IN WRITING THE COPYRIGHT HOLDERS AND/OR OTHER PARTIES
PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY OF ANY KIND, EITHER EXPRESSED
OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. THE ENTIRE RISK AS
TO THE QUALITY AND PERFORMANCE OF THE PROGRAM IS WITH YOU. SHOULD THE
PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF ALL NECESSARY SERVICING,
REPAIR OR CORRECTION.
12. IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MAY MODIFY AND/OR
REDISTRIBUTE THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES,
INCLUDING ANY GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING
OUT OF THE USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED
TO LOSS OF DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY
YOU OR THIRD PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER
PROGRAMS), EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE
POSSIBILITY OF SUCH DAMAGES.
END OF TERMS AND CONDITIONS
How to Apply These Terms to Your New Programs
If you develop a new program, and you want it to be of the greatest
possible use to the public, the best way to achieve this is to make it
free software which everyone can redistribute and change under these terms.
To do so, attach the following notices to the program. It is safest
to attach them to the start of each source file to most effectively
convey the exclusion of warranty; and each file should have at least
the "copyright" line and a pointer to where the full notice is found.
<one line to give the program's name and a brief idea of what it does.>
Copyright (C) <year> <name of author>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, see <https://www.gnu.org/licenses/>.
Also add information on how to contact you by electronic and paper mail.
If the program is interactive, make it output a short notice like this
when it starts in an interactive mode:
Gnomovision version 69, Copyright (C) year name of author
Gnomovision comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
This is free software, and you are welcome to redistribute it
under certain conditions; type `show c' for details.
The hypothetical commands `show w' and `show c' should show the appropriate
parts of the General Public License. Of course, the commands you use may
be called something other than `show w' and `show c'; they could even be
mouse-clicks or menu items--whatever suits your program.
You should also get your employer (if you work as a programmer) or your
school, if any, to sign a "copyright disclaimer" for the program, if
necessary. Here is a sample; alter the names:
Yoyodyne, Inc., hereby disclaims all copyright interest in the program
`Gnomovision' (which makes passes at compilers) written by James Hacker.
<signature of Moe Ghoul>, 1 April 1989
Moe Ghoul, President of Vice
This General Public License does not permit incorporating your program into
proprietary programs. If your program is a subroutine library, you may
consider it more useful to permit linking proprietary applications with the
library. If this is what you want to do, use the GNU Lesser General
Public License instead of this License.

37
Pipfile Normal file
View file

@ -0,0 +1,37 @@
[[source]]
url = "https://pypi.org/simple"
verify_ssl = true
name = "pypi"
[[source]]
url = "https://pypi.code.tiararodney.com/root/byteb4rb1e/+simple/"
verify_ssl = true
name = "pypicodetiararodney"
[packages]
"tiararodney.sekft" = {file = ".", editable = true}
"tiararodney.posix-sdc" = {version = "*", index = "pypicodetiararodney", extras= ["hub"]}
[dev-packages]
tox = "*"
pytest = "*"
mypy = "*"
build = "*"
twine = "*"
setuptools-scm = "~=8.2.0"
pypi-attestations = "*"
autopep8 = "*"
"tiararodney.posix-sdc" = {ref = "develop", git = "https://git.code.tiararodney.com/tiara/posix-sdc.git", extras = ["hub"]}
[requires]
python_version = "3"
[scripts]
"dist" = "python3 -m build"
"dist:attestations" = "python3 -m pypi_attestations sign dist/*"
"dist:publish:tiararodney" = "python3 -m twine upload --sign --repository tiararodney dist/*"
"test" = "tox"
"test:static" = "tox run -m static"
"test:unit" = "tox run -m unit"
"test:integration" = "tox run -m integration"
"test:smoke" = "tox run -m smoke"

1660
Pipfile.lock generated Normal file
View file

File diff suppressed because it is too large Load diff

79
README.md Normal file
View file

@ -0,0 +1,79 @@
# sekft
Fine-tune small open models to operate a POSIX shell as a self-directed citizen:
land with **no imperative**, discover where directives live, learn the provider
from its own self-documentation, do the work, and terminate (`exit` on success,
`panic` when genuinely blocked).
sekft is the **training half**. The dataset and the synthetic-data factory live
in [`posix-sdc`](../posix-sdc) (`tiararodney.posix-sdc`), which this package
depends on. Here live the trainer, the behavioural evaluator, and the
resident-base harness.
## Components
- **`sekft.sft`** (`sekft-train`) — supervised fine-tuner. Renders trajectories
with the tokenizer's own chat template and trains an **assistant-only** loss
mask (the commands plus the terminal token; environment turns masked to -100)
into a QLoRA adapter. Getting the mask wrong is the classic way to ruin a
shell-operator SFT, so it is the part tested hardest.
- **`sekft.eval`** (`sekft-eval`) — behavioural eval. Train loss says nothing
about whether the model operates the shell and leaves. This drops base +
adapter into held-out scenarios with no scaffold and reports the rates that
count: reach command-mode, terminate, checker passes.
- **`sekft.resident`** (`sekft-resident`) — resident-base harness. Loads the
14 GB base once and keeps it hot, training and evaluating adapters without
reloading it (over OcuLink/PCIe the base transfer otherwise dominates every
run).
## The render contract
The render the model trains on MUST equal what it is served with. The serving
harness (ccpty) sends structured `{role, content}` messages over the OpenAI
chat-completions protocol, so the endpoint applies the **model's own chat
template**. sekft therefore renders with `apply_chat_template`, after
`normalize_for_template` canonicalises each session: a leading `system` turn is
folded into the first `user` turn and consecutive same-role turns are merged,
because instruct templates such as Mistral's have no system role and require
strict user/assistant alternation. The same canonicalisation must run
serve-side, or train and serve diverge.
## Install
The training paths only run on a CUDA host, so the GPU stack is an extra:
```sh
pipenv install # editable sekft + the local editable posix-sdc
pipenv install -e '.[gpu]' # torch / transformers / peft / datasets, on the box
```
`pyproject.toml` declares `tiararodney.posix-sdc` abstractly; the `Pipfile`
overrides it with the local editable `../posix-sdc` for side-by-side development.
## Use (on the GPU box)
```sh
# fine-tune an adapter on the posix-sdc trajectories
sekft-train --data ./trajectories --base mistralai/Mistral-7B-Instruct-v0.2 \
--out ./ckpt --load-4bit
# inspect the assistant-only loss mask without training (runs anywhere)
sekft-train --data ./trajectories --base <dir> --inspect
# behavioural eval on held-out scenario bundles (worlds, not trajectories)
sekft-eval --base <dir> --adapter ./ckpt --scenarios ./holdout --n 16
# resident loop: load the base once, cycle adapters without reloading it
sekft-resident --base <dir> --load-4bit
```
The eval consumes held-out **scenario bundles** from posix-sdc (it stands up and
verifies each in a fresh container), not trajectories.
## Result
Fine-tuning `mistralai/Mistral-7B-Instruct-v0.2` on the posix-sdc data lifted
clean termination on archetype-level held-out scenarios from **0/16 (base) to
9/16 (tuned)**: the operate-and-terminate mechanism generalised to unseen task
types, while task competence stayed archetype-local. See the experiment
[*From seed to weights*](https://blog.tiararodney.com/projects/2026/semantic-execution-kernel/experiments/from-seed-to-weights/).

234
TODO
View file

@ -15,3 +15,237 @@ Mappings:
- Module: sekft - Module: sekft
Product: sek Product: sek
Component: sekft Component: sekft
--ISSUE
Content-Type: application/issue
ID: 1
Type: feature
Title: Package sekft as an installable namespace package
Status: done
Priority: medium
Created: 2026-06-16
Module: sekft
Relationships:
Description: Turn the flat trainer scripts into an installable tiararodney.sekft
namespace package: src layout, pyproject with the abstract
posix-sdc dependency and an optional gpu extra, console scripts, a
Pipfile pinning posix-sdc as a local editable override, and tox
environments.
--ISSUE
Content-Type: application/issue
ID: 2
Type: feature
Title: SFT trainer with chat-template render and assistant-only mask
Status: done
Priority: medium
Created: 2026-06-16
Module: sekft
Relationships:
Description: Add the supervised fine-tuner: render trajectories through the
tokenizer's own chat template (matching serving), canonicalise
turns (fold system, merge consecutive), derive an assistant-only
loss mask by token-prefix differencing, and train a QLoRA adapter.
--ISSUE
Content-Type: application/issue
ID: 3
Type: feature
Title: Behavioural evaluator
Status: done
Priority: medium
Created: 2026-06-16
Module: sekft
Relationships:
Description: Add the behavioural eval: load base plus LoRA adapter, drop it into
held-out scenarios with no scaffold, drive them through a local
operator that renders with the model's chat template, and report
reach/terminate/checker rates.
--ISSUE
Content-Type: application/issue
ID: 4
Type: feature
Title: Resident-base train/eval harness
Status: done
Priority: medium
Created: 2026-06-16
Module: sekft
Relationships:
Description: Add the resident harness that loads the 14GB base once and keeps it
hot, training fresh LoRA adapters and evaluating them without
reloading the base, for the slow-OcuLink iterate loop.
--ISSUE
Content-Type: application/issue
ID: 5
Type: feature
Title: Pipeline overview README
Status: done
Priority: medium
Created: 2026-06-16
Module: sekft
Relationships:
Description: Document the sekft pipeline: the trainer, evaluator, and resident
harness; how they consume the posix-sdc dataset; the render
contract; and how to run on the GPU box.
--ISSUE
Content-Type: application/issue
ID: 6
Type: feature
Title: Test suite: unit and smoke
Status: done
Priority: medium
Created: 2026-06-16
Module: sekft
Relationships:
Description: Add a pytest suite: torch-free unit tests for the render
canonicalisation and assistant-only mask (fake tokenizer), and
smoke tests that the console entry points respond to --help without
the GPU stack.
--ISSUE
Content-Type: application/issue
ID: 7
Type: feature
Title: Add GPL-2.0 license and drop the relocated Dockerfile
Status: done
Priority: medium
Created: 2026-06-16
Module: sekft
Relationships:
Description: License sekft under GPL-2.0 (canonical text plus pyproject
metadata) and remove the dash Dockerfile, which now lives in
posix-sdc under docker/alpine-dash.
--ISSUE
Content-Type: application/issue
ID: 8
Type: feature
Title: Refresh docs for the packaged trainer
Status: done
Priority: medium
Created: 2026-06-16
Module: sekft
Relationships:
Description: The README still describes sekft as the data factory
(generate/rollout/dashdocker/taxonomy/schema), which all moved to
posix-sdc. Rewrite it as the trainer (sft/eval/resident) that
consumes posix-sdc, and update the module docstrings to
console-script invocations and the chat-template render contract.
--ISSUE
Content-Type: application/issue
ID: 9
Type: feature
Title: Type-check the package under mypy strict
Status: done
Priority: medium
Created: 2026-06-17
Module: sekft
Relationships:
Description: Make the lint env honestly pass: add mypy as a dev dependency,
ignore_missing_imports for the ML libs, fully annotate
eval/resident/sft (including the inner operator callables), and
ship a py.typed marker so the Typing::Typed claim is real.
--ISSUE
Content-Type: application/issue
ID: 10
Type: feature
Title: structured logging for the trainer (sft)
Status: done
Priority: medium
Created: 2026-06-17
Module: sekft
Relationships:
Description: The trainer is nearly silent: outside an example count and a save
line it prints nothing through tokenizer load, the ~14GB base-model
load, example building, and the whole training loop, and
trajectories dropped for exceeding --max-len or having an empty
loss mask vanish without a trace. Add a small shared logging setup
(_log.py, stderr so stdout stays clean for results) and a module
logger; give sekft-train -v/--verbose and -q/--quiet. Log the run
config and each phase, report dataset accounting (keepers ->
usable, with counts dropped for length / empty-mask and a warning
when any are dropped), and raise transformers' verbosity during
training so the per-step curve shows. Apply to train() and
inspect().
--ISSUE
Content-Type: application/issue
ID: 11
Type: bugfix
Title: operate_rate can sum a None (eval + resident)
Status: done
Priority: medium
Created: 2026-06-17
Module: sekft
Relationships:
Description: operate_rate computes sum(t.steps > 0 and t.meta.get('clean') for t
in rows). The 'and' yields the right operand when steps>0, so if
meta lacks the 'clean' key it yields None and sum() raises
TypeError at runtime; mypy (now that posix-sdc ships py.typed and
Trajectory is typed) flags the generator item type in eval.py:83
and resident.py:157. Wrap the predicate in bool() so it counts
trajectories that operated and are clean, fixing both the type
error and the latent crash.
--ISSUE
Content-Type: application/issue
ID: 12
Type: feature
Title: load training data from a raw dir, a curated jsonl, or the Hub
Status: done
Priority: medium
Created: 2026-06-17
Module: sekft
Relationships:
Description: iter_keepers reads only raw per-trajectory .json - one of three
input shapes the trainer should accept. Add load_turns(data, hub,
revision) that yields assistant-bearing turns from: a directory of
raw rollout .json (keep-filtered, today's iter_keepers); a curated
.jsonl corpus file (already keep-filtered, yield turns per line);
or the published corpus via posix-sdc's load_trajectories (local
data/ in a checkout, else the Hub). sekft-train gains --hub and
--revision; --data dispatches by dir-vs-.jsonl. Raw-rollout reading
stays sekft-local; curated+Hub reuse posix-sdc's loader (imported
lazily so the trainer needs neither posix-sdc nor huggingface_hub
for the raw/jsonl paths). Unit tests for the raw-dir and jsonl
dispatch.
--ISSUE
Content-Type: application/issue
ID: 13
Type: feature
Title: reference posix-sdc three ways for seamless multi-machine dev
Status: done
Priority: medium
Created: 2026-06-17
Module: sekft
Relationships:
Description: Wire the posix-sdc dependency as a triplet: the abstract
posix-sdc[hub] in pyproject (so the trainer's --hub path can reach
the Hub via huggingface_hub); the published wheel from the private
index in Pipfile [packages]; the git develop branch in Pipfile
[dev-packages] for develop-time. Commit Pipfile.lock so the
dependency surface and lock land together.
--ISSUE
Content-Type: application/issue
ID: 14
Type: bugfix
Title: refresh Pipfile.lock against published posix-sdc 1.2.2
Status: done
Priority: medium
Created: 2026-06-17
Module: sekft
Relationships:
Description: The lock committed with the triplet (#13) predated the published
posix-sdc 1.2.2 wheel, so it could not pin the real [hub] closure.
Now that 1.2.2 is on the private index, re-lock: posix-sdc resolves
to ==1.2.2 from the index and the [hub] extra pulls huggingface_hub
and its transitive deps into the lock. Commit the refreshed
Pipfile.lock so the next machine installs the published wheel with
the Hub path available.

92
pyproject.toml Normal file
View file

@ -0,0 +1,92 @@
[build-system]
requires = [
"setuptools",
"wheel",
"setuptools-scm[toml]"
]
build-backend = "setuptools.build_meta"
[project]
name = "tiararodney.sekft"
description = "Fine-tune small open models to operate a POSIX shell (sek)"
authors = [
{ name = "Tiara Rodney", email = "tiara.rodney@byteb4rb1e.me" }
]
license-files = ["LICENSE"]
readme = "README.md"
classifiers = [
"Development Status :: 3 - Alpha",
"Intended Audience :: Developers",
"Intended Audience :: Science/Research",
"License :: OSI Approved :: GNU General Public License v2 (GPLv2)",
"Natural Language :: English",
"Operating System :: POSIX :: Linux",
"Programming Language :: Python :: 3",
"Programming Language :: Python :: 3.9",
"Programming Language :: Python :: 3.10",
"Programming Language :: Python :: 3.11",
"Programming Language :: Python :: 3.12",
"Topic :: Scientific/Engineering :: Artificial Intelligence",
"Topic :: System :: Shells",
"Typing :: Typed",
]
dependencies = [
"tiararodney.posix-sdc[hub]",
]
dynamic = ["version"]
requires-python = ">=3.9"
[project.optional-dependencies]
gpu = [
"torch",
"transformers",
"peft",
"datasets",
"accelerate",
"bitsandbytes",
"tensorboard",
]
[project.scripts]
sekft-train = "tiararodney.sekft.sft:main"
sekft-eval = "tiararodney.sekft.eval:main"
sekft-resident = "tiararodney.sekft.resident:main"
[project.urls]
Git = "https://git.code.tiararodney.com/tiararodney/sekft"
[tool.setuptools.packages.find]
where = ["src"]
namespaces = true
[tool.setuptools.package-data]
"tiararodney.sekft" = ["py.typed"]
[tool.pytest.ini_options]
pythonpath = ["src", "../posix-sdc/src"]
testpaths = ["tests"]
markers = [
"pytest: integration tests runnable without external services",
"gpu: requires torch and a GPU",
"docker: requires Docker and the sekft-dash image",
]
[tool.mypy]
strict = true
mypy_path = "src"
explicit_package_bases = true
namespace_packages = true
[[tool.mypy.overrides]]
module = [
"torch.*", "transformers.*", "peft.*", "datasets.*", "bitsandbytes.*",
"tiararodney.posix_sdc.*",
]
ignore_missing_imports = true
[tool.autopep8]
max_line_length = 80
aggressive = 3
recursive = true
[tool.setuptools_scm]

5
src/tiararodney/sekft/__init__.py Normal file
View file

@ -0,0 +1,5 @@
"""sekft: fine-tune small open models to operate a POSIX shell (sek).
Consumes the posix-sdc dataset; the trainer, behavioural evaluator, and the
resident-base harness live here.
"""

20
src/tiararodney/sekft/_log.py Normal file
View file

@ -0,0 +1,20 @@
"""Console logging setup shared by the sekft entry points.
Logs go to stderr so stdout stays clean for a command's actual output (metrics
JSON, a path a caller might capture). Call :func:`setup` once at the top of a
``main()``; modules then log through ``logging.getLogger("sekft.<area>")``.
"""
from __future__ import annotations
import logging
def setup(verbose: bool = False, quiet: bool = False) -> None:
"""Configure root logging to stderr. ``quiet`` shows warnings and worse,
``verbose`` adds debug; the default is info."""
level = logging.WARNING if quiet else logging.DEBUG if verbose else logging.INFO
logging.basicConfig(
level=level,
format="%(asctime)s %(levelname)-5s %(name)s %(message)s",
datefmt="%H:%M:%S",
)

105
src/tiararodney/sekft/eval.py Normal file
View file

@ -0,0 +1,105 @@
"""Behavioural eval: the metric that matters.
Train loss says nothing about whether the model operates the shell and leaves.
This loads a fine-tuned model (base + LoRA adapter), drops it into held-out
scenarios with NO scaffold (the trained behaviour must stand on its own), and
reports the rates that count: does it reach command-mode, does it terminate,
does the checker pass.
sekft-eval --base <hf-dir> --adapter ./ckpt-mistral-r16 \
--scenarios ./holdout-scenarios --n 10
Reuses the posix-sdc rollout loop with a *local* operator: the model renders and
generates with the same chat template it was trained on (train == eval == serve,
via ``apply_chat_template`` + ``normalize_for_template``, or the prompts go out
of distribution). Prerequisites on the box: torch + transformers + peft, the
``sekft-dash`` image, and held-out SCENARIO bundles from the posix-sdc factory
(not trajectories; the eval stands up and verifies each).
"""
from __future__ import annotations
import argparse
import json
from collections.abc import Callable
from pathlib import Path
from typing import Any
from tiararodney.posix_sdc.factory.dashdocker import DashDocker, available
from tiararodney.posix_sdc.factory.rollout import rollout
from tiararodney.posix_sdc.schema import Scenario
from .sft import normalize_for_template
def make_local_operator(base: str, adapter: str, max_new_tokens: int = 64,
temperature: float = 0.7) -> Callable[[list[dict[str, str]]], str]:
"""A ``messages -> command`` callable backed by base + LoRA adapter.
Renders the conversation exactly as the model was trained, appends the
assistant header, generates one turn, and cuts at the first stop marker.
"""
import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained(adapter)
model = AutoModelForCausalLM.from_pretrained(
base, torch_dtype=torch.float16, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)
model.eval()
def operator(messages: list[dict[str, str]]) -> str:
msgs = normalize_for_template(messages)
ids = tok.apply_chat_template(
msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
with torch.no_grad():
out = model.generate(
ids, max_new_tokens=max_new_tokens,
do_sample=temperature > 0, temperature=max(temperature, 1e-2),
eos_token_id=tok.eos_token_id, pad_token_id=tok.eos_token_id)
text: str = tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True).strip()
return text
return operator
def evaluate(base: str, adapter: str, scenarios_dir: Path, n: int,
max_steps: int, temperature: float) -> dict[str, Any]:
if not available():
raise SystemExit("sekft-dash image unavailable; `docker build -t sekft-dash .`")
operator = make_local_operator(base, adapter, temperature=temperature)
backend = DashDocker()
rows = []
for f in sorted(scenarios_dir.glob("*.json"))[:n]:
sc = Scenario.from_dict(json.loads(f.read_text()))
tj = rollout(sc, backend, max_steps=max_steps, temperature=temperature,
operator=operator, use_scaffold=False)
rows.append(tj)
print(f" {sc.id}: {tj.outcome} (terminal={tj.terminal} "
f"verified={tj.verified} steps={tj.steps})")
d = len(rows) or 1
return {
"n": len(rows),
"operate_rate": round(sum(bool(t.steps > 0 and t.meta.get("clean")) for t in rows) / d, 3),
"terminate_rate": round(sum(t.terminal in ("exit", "panic") for t in rows) / d, 3),
"verified_rate": round(sum(t.verified for t in rows) / d, 3),
"clean_rate": round(sum(t.keep for t in rows) / d, 3),
}
def main() -> None:
ap = argparse.ArgumentParser(description="Behavioural eval of a tuned model.")
ap.add_argument("--base", required=True)
ap.add_argument("--adapter", required=True)
ap.add_argument("--scenarios", type=Path, required=True)
ap.add_argument("--n", type=int, default=10)
ap.add_argument("--max-steps", type=int, default=30)
ap.add_argument("--temperature", type=float, default=0.7)
ns = ap.parse_args()
m = evaluate(ns.base, ns.adapter, ns.scenarios, ns.n, ns.max_steps, ns.temperature)
print("\n=== behavioural metrics ===")
print(json.dumps(m, indent=2))
if __name__ == "__main__":
main()

0
src/tiararodney/sekft/py.typed Normal file
View file

189
src/tiararodney/sekft/resident.py Normal file
View file

@ -0,0 +1,189 @@
"""Resident harness: load the base ONCE, cycle adapters.
On a slow link (OcuLink / PCIe 3.0 x4) the 14 GB base transfer dominates every
process start. This loads the base once and keeps it hot, so the
iterate-train-eval loop pays the transfer only at startup. Each ``fit`` trains a
fresh LoRA adapter on the resident base and ``unload``s it back to clean; each
``evaluate`` attaches a saved adapter for inference and unloads.
Interactive (IPython on the GPU box) is the intended use:
from tiararodney.sekft.resident import Resident
r = Resident("~/llm-models/mistral-7b-instruct-v0.2", load_4bit=True)
r.fit("~/sekft/trajectories", "~/sekft/ckpt-a", lora_r=16, lr=2e-4, epochs=3)
r.evaluate("~/sekft/ckpt-a", "~/sekft/holdout", n=10)
r.fit("~/sekft/trajectories", "~/sekft/ckpt-b", lora_r=32) # NO base reload
Or `sekft-resident --base <dir> --selftest-data <stub_dir>` to prove the base
loads once and two adapters train against it.
"""
from __future__ import annotations
import argparse
import gc
import json
from pathlib import Path
from typing import Any
import torch
from datasets import Dataset
from peft import (LoraConfig, PeftModel, get_peft_model,
prepare_model_for_kbit_training)
from transformers import (AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig,
DataCollatorForSeq2Seq, Trainer, TrainingArguments)
from .sft import build_masked_example, iter_keepers, normalize_for_template
LORA_TARGETS = ["q_proj", "k_proj", "v_proj", "o_proj"]
def _free() -> None:
gc.collect()
torch.cuda.empty_cache()
class Resident:
"""A base model held resident on the GPU; adapters cycle through it."""
def __init__(self, base: str, load_4bit: bool = False) -> None:
self.base_path = str(Path(base).expanduser())
self.load_4bit = load_4bit
self.tok = AutoTokenizer.from_pretrained(self.base_path)
if self.tok.pad_token is None:
self.tok.pad_token = self.tok.eos_token
quant = None
if load_4bit:
quant = BitsAndBytesConfig(
load_in_4bit=True, bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.float16, bnb_4bit_use_double_quant=True)
print(f"[resident] loading base ONCE: {self.base_path} (4bit={load_4bit}) ...")
self.base = AutoModelForCausalLM.from_pretrained(
self.base_path, dtype=torch.float16, quantization_config=quant)
self.base = (prepare_model_for_kbit_training(self.base) if load_4bit
else self.base)
if not load_4bit:
self.base.enable_input_require_grads()
dev = next(self.base.parameters()).device
mem = torch.cuda.memory_allocated() / 1e9
print(f"[resident] base resident on {dev}; {mem:.1f} GB VRAM")
# -- build masked rows from kept trajectories --------------------------
def _rows(self, data_dir: Path, max_len: int) -> list[dict[str, list[Any]]]:
rows = []
for turns in iter_keepers(data_dir):
ex = build_masked_example(turns, self.tok)
if len(ex["input_ids"]) <= max_len and any(l != -100 for l in ex["labels"]):
rows.append(ex)
if not rows:
raise SystemExit(f"no usable keeper trajectories in {data_dir}")
return rows
# -- train a fresh adapter on the resident base ------------------------
def fit(self, data_dir: str, out: str, lora_r: int = 16, lr: float = 2e-4,
epochs: float = 3.0, batch: int = 1, accum: int = 8,
max_len: int = 4096) -> Path:
ddir, odir = Path(data_dir).expanduser(), Path(out).expanduser()
ds = Dataset.from_list(self._rows(ddir, max_len))
if not self.load_4bit:
self.base.gradient_checkpointing_enable()
model = get_peft_model(self.base, LoraConfig(
r=lora_r, lora_alpha=lora_r * 2, lora_dropout=0.05,
task_type="CAUSAL_LM", target_modules=LORA_TARGETS))
model.print_trainable_parameters()
args = TrainingArguments(
output_dir=str(odir), per_device_train_batch_size=batch,
gradient_accumulation_steps=accum, num_train_epochs=epochs,
learning_rate=lr, fp16=True, logging_steps=1, save_strategy="no",
report_to=["tensorboard"], logging_dir=str(odir / "runs"),
remove_unused_columns=False, warmup_ratio=0.03)
tr = Trainer(model=model, args=args, train_dataset=ds,
data_collator=DataCollatorForSeq2Seq(
self.tok, padding=True, label_pad_token_id=-100))
tr.train()
odir.mkdir(parents=True, exist_ok=True)
model.save_pretrained(str(odir))
self.tok.save_pretrained(str(odir))
(odir / "log_history.jsonl").write_text(
"\n".join(json.dumps(r) for r in tr.state.log_history))
losses = [h["loss"] for h in tr.state.log_history if "loss" in h]
print(f"[resident] fit -> {odir} final loss {losses[-1] if losses else '?'}")
self.base = model.unload() # strip LoRA, restore resident base
del model, tr, ds
_free()
return odir
# -- behavioural eval of a saved adapter -------------------------------
def evaluate(self, adapter: str, scenarios_dir: str, n: int = 10,
max_steps: int = 30, temperature: float = 0.7) -> dict[str, Any]:
from tiararodney.posix_sdc.factory.dashdocker import DashDocker, available
from tiararodney.posix_sdc.factory.rollout import rollout
from tiararodney.posix_sdc.schema import Scenario
if not available():
raise SystemExit("sekft-dash image unavailable on this box")
# adapter=None -> evaluate the BASE model (the within-holdout baseline).
if adapter:
adapter = str(Path(adapter).expanduser())
pm = PeftModel.from_pretrained(self.base, adapter)
else:
pm = self.base
pm.eval()
def operator(messages: list[dict[str, str]]) -> str:
msgs = normalize_for_template(messages)
ids = self.tok.apply_chat_template(
msgs, add_generation_prompt=True, return_tensors="pt").to(pm.device)
with torch.no_grad():
o = pm.generate(ids, max_new_tokens=64, do_sample=temperature > 0,
temperature=max(temperature, 1e-2),
eos_token_id=self.tok.eos_token_id,
pad_token_id=self.tok.eos_token_id)
text: str = self.tok.decode(o[0][ids.shape[1]:], skip_special_tokens=True).strip()
return text
backend = DashDocker()
rows = []
for f in sorted(Path(scenarios_dir).expanduser().glob("*.json"))[:n]:
sc = Scenario.from_dict(json.loads(f.read_text()))
tj = rollout(sc, backend, max_steps=max_steps, temperature=temperature,
operator=operator, use_scaffold=False)
rows.append(tj)
print(f" {sc.id}: {tj.outcome} terminal={tj.terminal} verified={tj.verified}")
d = len(rows) or 1
m = {
"n": len(rows),
"operate_rate": round(sum(bool(t.steps > 0 and t.meta.get("clean")) for t in rows) / d, 3),
"terminate_rate": round(sum(t.terminal in ("exit", "panic") for t in rows) / d, 3),
"verified_rate": round(sum(t.verified for t in rows) / d, 3),
"clean_rate": round(sum(t.keep for t in rows) / d, 3),
}
if adapter: # base is unwrapped only if we wrapped it
self.base = pm.unload()
del pm
_free()
print("[resident] eval:", json.dumps(m))
return m
def main() -> None:
ap = argparse.ArgumentParser(description="Resident base; cycle adapters.")
ap.add_argument("--base", required=True)
ap.add_argument("--load-4bit", action="store_true")
ap.add_argument("--selftest-data",
help="fit two adapters on this data to prove resident multi-fit")
ns = ap.parse_args()
r = Resident(ns.base, ns.load_4bit)
if ns.selftest_data:
print("=== selftest: two fits on the SAME resident base (no reload) ===")
r.fit(ns.selftest_data, "/tmp/res-a", epochs=1, lora_r=8)
r.fit(ns.selftest_data, "/tmp/res-b", epochs=1, lora_r=8)
print("=== selftest OK: base loaded once, two adapters trained ===")
else:
print("Resident ready. Import and use r.fit() / r.evaluate(), "
"or pass --selftest-data <dir>.")
if __name__ == "__main__":
main()

289
src/tiararodney/sekft/sft.py Normal file
View file

@ -0,0 +1,289 @@
"""sekft trainer: SFT a base model on kept shell-operation trajectories.
Trains assistant turns ONLY -- the commands and the terminal ``exit`` / ``panic``.
The environment turns (system orientation, prompts, command output) are masked
to ``-100`` so the model learns to *produce* commands, not to predict the
environment's replies. Getting this mask wrong is the classic way to ruin a
shell-operator SFT (the model starts hallucinating output), so it is the part
worth testing hardest -- and it is framework-independent.
Render uses the tokenizer's OWN chat template (``apply_chat_template``), so the
training render is identical to what the serving harness produces (ccpty sends
structured messages and the inference endpoint applies the model's default
template). Trajectories are canonicalised first (``normalize_for_template``):
a leading ``system`` turn is folded into the first ``user`` turn and consecutive
same-role turns are merged, because instruct templates such as Mistral's have no
system role and require strict user/assistant alternation. That same
canonicalisation must run on the serving side. Everything else is standard
causal-LM SFT with an assistant-only loss mask.
sekft-train --data ./trajectories --base <hf-model-dir> --out ./ckpt
sekft-train --data corpus.jsonl --base <dir> # a curated .jsonl corpus
sekft-train --hub --base <dir> # the published corpus (Hub)
sekft-train --data ./trajectories --base <dir> --inspect # mask stats, no training
Training needs torch + transformers + peft (a GPU box). ``--inspect`` and the
normalize/mask helpers run anywhere a tokenizer with a chat template is
available.
"""
from __future__ import annotations
import argparse
import json
import logging
from collections.abc import Iterator
from pathlib import Path
from typing import Any
from ._log import setup as _setup_logging
log = logging.getLogger("sekft.train")
def normalize_for_template(messages: list[dict[str, str]]) -> list[dict[str, str]]:
"""Canonicalise a trajectory for instruct chat templates that have no system
role and require strict user/assistant alternation (Mistral and friends):
treat ``system`` as ``user``, then merge consecutive same-role turns by
joining their content with a newline.
This is loss-neutral for the assistant mask (only environment/user turns
ever merge; the assistant commands are never adjacent in this data) and it
is what lets ``apply_chat_template`` render the multi-turn shell dialogue.
The serving side MUST apply the same canonicalisation, or train and serve
diverge again.
"""
out: list[dict[str, str]] = []
for m in messages:
role = "user" if m["role"] == "system" else m["role"]
if out and out[-1]["role"] == role:
out[-1] = {"role": role, "content": out[-1]["content"] + "\n" + m["content"]}
else:
out.append({"role": role, "content": m["content"]})
return out
def build_masked_example(messages: list[dict[str, str]], tokenizer: Any) -> dict[str, list[Any]]:
"""Tokenize a trajectory with the tokenizer's OWN chat template and build an
assistant-only loss mask.
The render is ``tokenizer.apply_chat_template`` on the canonicalised turns,
so it is byte-identical to what the serving harness sends. The mask is
derived by token-prefix differencing: the tokens an assistant turn
contributes are exactly those that appear when it extends the rendered
prefix, which trains the commands plus the template's end-of-turn token (so
the model learns to stop) and masks every environment turn to ``-100``. This
assumes an additive template (each turn extends the previous render); a
non-additive one raises rather than silently mis-mask.
"""
msgs = normalize_for_template(messages)
ids = tokenizer.apply_chat_template(msgs, add_generation_prompt=False)
labels = [-100] * len(ids)
prev: list[int] = []
for i, m in enumerate(msgs):
upto = tokenizer.apply_chat_template(msgs[:i + 1], add_generation_prompt=False)
if ids[:len(upto)] != upto or upto[:len(prev)] != prev:
raise ValueError("chat template is not additive; cannot derive an "
"assistant loss mask by token-prefix differencing")
if m["role"] == "assistant":
for j in range(len(prev), len(upto)):
labels[j] = ids[j]
prev = upto
return {"input_ids": ids, "attention_mask": [1] * len(ids), "labels": labels}
def iter_keepers(data_dir: Path) -> Iterator[list[dict[str, str]]]:
"""Yield ``turns`` (message lists) from raw rollout JSONs marked keep."""
for f in sorted(data_dir.glob("*.json")):
d = json.loads(f.read_text())
if d.get("keep"):
yield d["turns"]
def load_turns(data: Path, hub: bool = False,
revision: str | None = None) -> Iterator[list[dict[str, str]]]:
"""Yield assistant-bearing ``turns`` from one of three sources:
- ``--hub``: the published corpus via posix-sdc's ``load_trajectories`` (the
in-repo ``data/`` of a posix-sdc checkout, else the Hugging Face Hub);
- ``data`` a ``.jsonl`` file: a curated corpus, already keep-filtered, one
record per line;
- ``data`` a directory: raw rollout ``.json`` (keep-filtered here).
posix-sdc is imported lazily, so the raw-dir and ``.jsonl`` paths need
neither posix-sdc nor huggingface_hub installed.
"""
if hub:
from tiararodney.posix_sdc import load_trajectories
for r in load_trajectories(revision=revision):
yield r["turns"]
elif data.is_dir():
yield from iter_keepers(data)
elif data.suffix == ".jsonl":
with open(data) as fh:
for line in fh:
if line.strip():
yield json.loads(line)["turns"]
else:
raise SystemExit(
f"--data must be a rollout directory or a .jsonl corpus (got {data})")
def mask_stats(example: dict[str, list[Any]]) -> tuple[int, int]:
"""(trained tokens, total tokens) for an example."""
trained = sum(1 for x in example["labels"] if x != -100)
return trained, len(example["labels"])
# --------------------------------------------------------------------------
# Training (GPU box: torch + transformers + peft)
# --------------------------------------------------------------------------
def train(data_dir: Path, base: str, out: Path, epochs: float, lr: float,
batch: int, accum: int, max_len: int, lora_r: int,
load_4bit: bool = False, hub: bool = False,
revision: str | None = None) -> None:
import torch
from datasets import Dataset
from peft import LoraConfig, get_peft_model
from transformers import (AutoModelForCausalLM, AutoTokenizer,
DataCollatorForSeq2Seq, Trainer, TrainingArguments)
from transformers.utils import logging as hf_logging
# Surface the Trainer's own per-step curve (loss/lr/grad_norm); it is at
# WARNING by default, which is most of why training looks silent.
hf_logging.set_verbosity_info()
source = "hub" if hub else data_dir
log.info("base=%s data=%s out=%s", base, source, out)
log.info("loading tokenizer: %s", base)
tok = AutoTokenizer.from_pretrained(base)
if tok.pad_token is None:
tok.pad_token = tok.eos_token
log.info("building masked examples from %s ...", source)
rows: list[dict[str, list[Any]]] = []
n_seen = n_long = n_empty = 0
for turns in load_turns(data_dir, hub=hub, revision=revision):
n_seen += 1
ex = build_masked_example(turns, tok)
log.debug(" trajectory %d: %d turns -> %d tokens, %d trained",
n_seen, len(turns), len(ex["input_ids"]), mask_stats(ex)[0])
if n_seen % 100 == 0:
log.info(" ... %d trajectories processed, %d usable", n_seen, len(rows))
if len(ex["input_ids"]) > max_len:
n_long += 1
continue
if not any(l != -100 for l in ex["labels"]):
n_empty += 1
continue
rows.append(ex)
if not rows:
raise SystemExit(f"no usable keeper trajectories in {data_dir}")
trained = sum(mask_stats(r)[0] for r in rows)
total = sum(mask_stats(r)[1] for r in rows)
log.info("dataset: %d keepers -> %d usable; %d trained / %d tokens (%.1f%% assistant)",
n_seen, len(rows), trained, total, 100 * trained / total)
if n_long or n_empty:
log.warning("dropped %d trajectories: %d over --max-len %d, %d empty-mask",
n_long + n_empty, n_long, max_len, n_empty)
ds = Dataset.from_list(rows)
# 4-bit (QLoRA) shrinks the base from ~14 GB to ~4 GB to move across the
# OcuLink/PCIe link and to hold in VRAM; nf4 + fp16 compute works on the
# V100 (sm_70). Without it, plain fp16 weights.
quant = None
if load_4bit:
from transformers import BitsAndBytesConfig
quant = BitsAndBytesConfig(
load_in_4bit=True, bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.float16, bnb_4bit_use_double_quant=True,
)
log.info("loading base model: %s (%s)", base,
"4-bit QLoRA" if load_4bit else "fp16")
model = AutoModelForCausalLM.from_pretrained(
base, dtype=torch.float16, quantization_config=quant)
if load_4bit:
from peft import prepare_model_for_kbit_training
model = prepare_model_for_kbit_training(model) # handles ckpt + input grads
else:
model.enable_input_require_grads()
model.gradient_checkpointing_enable()
model = get_peft_model(model, LoraConfig(
r=lora_r, lora_alpha=lora_r * 2, lora_dropout=0.05, task_type="CAUSAL_LM",
target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
))
n_train, n_all = model.get_nb_trainable_parameters()
log.info("LoRA r=%d: %d trainable / %d params (%.3f%%)",
lora_r, n_train, n_all, 100 * n_train / n_all)
args = TrainingArguments(
output_dir=str(out), per_device_train_batch_size=batch,
gradient_accumulation_steps=accum, num_train_epochs=epochs,
learning_rate=lr, fp16=True, logging_steps=1, save_strategy="epoch",
report_to=["tensorboard"], logging_dir=str(out / "runs"),
remove_unused_columns=False, warmup_ratio=0.03,
)
trainer = Trainer(
model=model, args=args, train_dataset=ds,
data_collator=DataCollatorForSeq2Seq(tok, padding=True, label_pad_token_id=-100),
)
log.info("training: %g epochs, lr=%g, batch=%d x accum=%d (effective %d), max_len=%d",
epochs, lr, batch, accum, batch * accum, max_len)
trainer.train()
model.save_pretrained(str(out))
tok.save_pretrained(str(out))
# durable, greppable record of the curve (loss/lr/grad_norm per step).
(out / "log_history.jsonl").write_text(
"\n".join(json.dumps(r) for r in trainer.state.log_history))
log.info("saved LoRA adapter + log_history.jsonl -> %s (tensorboard: --logdir %s)",
out, out / "runs")
def inspect(data_dir: Path, base: str, hub: bool = False,
revision: str | None = None) -> None:
from transformers import AutoTokenizer
log.info("loading tokenizer: %s", base)
tok = AutoTokenizer.from_pretrained(base)
n = tt = tr = 0
for turns in load_turns(data_dir, hub=hub, revision=revision):
ex = build_masked_example(turns, tok)
t, total = mask_stats(ex)
tr += t; tt += total; n += 1
if not n:
raise SystemExit(f"no keeper trajectories in {data_dir}")
log.info("%d keeper trajectories; %d/%d tokens trained (%.1f%% assistant, rest masked)",
n, tr, tt, 100 * tr / tt)
def main() -> None:
ap = argparse.ArgumentParser(description="SFT a model on shell trajectories.")
ap.add_argument("--data", type=Path, default=Path("./trajectories"),
help="a raw rollout dir or a curated .jsonl corpus")
ap.add_argument("--hub", action="store_true",
help="load the published corpus via posix-sdc (Hub); ignores --data")
ap.add_argument("--revision", default=None,
help="dataset revision/tag to pin when using --hub")
ap.add_argument("--base", required=True, help="HF model id or local dir")
ap.add_argument("--out", type=Path, default=Path("./ckpt"))
ap.add_argument("--inspect", action="store_true", help="mask stats only, no training")
ap.add_argument("--epochs", type=float, default=3.0)
ap.add_argument("--lr", type=float, default=2e-4)
ap.add_argument("--batch", type=int, default=1)
ap.add_argument("--accum", type=int, default=8)
ap.add_argument("--max-len", type=int, default=4096)
ap.add_argument("--lora-r", type=int, default=16)
ap.add_argument("--load-4bit", action="store_true",
help="QLoRA: load base in 4-bit (less to move over the link, less VRAM)")
ap.add_argument("-v", "--verbose", action="store_true", help="debug-level logging")
ap.add_argument("-q", "--quiet", action="store_true", help="warnings and errors only")
ns = ap.parse_args()
_setup_logging(verbose=ns.verbose, quiet=ns.quiet)
if ns.inspect:
inspect(ns.data, ns.base, hub=ns.hub, revision=ns.revision)
else:
train(ns.data, ns.base, ns.out, ns.epochs, ns.lr, ns.batch, ns.accum,
ns.max_len, ns.lora_r, ns.load_4bit, hub=ns.hub, revision=ns.revision)
if __name__ == "__main__":
main()

30
tests/smoke/test_entrypoints.py Normal file
View file

@ -0,0 +1,30 @@
"""Smoke tests: the console entry points load and respond to --help without the
GPU stack (torch is imported lazily inside the training/eval code paths)."""
from __future__ import annotations
import os
import subprocess
import sys
from pathlib import Path
ROOT = Path(__file__).resolve().parents[2]
SRC = ROOT / "src"
POSIX_SRC = ROOT.parent / "posix-sdc" / "src"
def _help(module: str) -> "subprocess.CompletedProcess[str]":
env = dict(os.environ, PYTHONPATH=os.pathsep.join([str(SRC), str(POSIX_SRC)]))
return subprocess.run([sys.executable, "-m", module, "--help"],
capture_output=True, text=True, env=env)
def test_train_help() -> None:
cp = _help("tiararodney.sekft.sft")
assert cp.returncode == 0, cp.stderr
assert "--data" in cp.stdout
def test_eval_help() -> None:
cp = _help("tiararodney.sekft.eval")
assert cp.returncode == 0, cp.stderr
assert "--adapter" in cp.stdout

35
tests/unit/test_load.py Normal file
View file

@ -0,0 +1,35 @@
"""Unit tests for the trainer's three-source data loader (raw dir / curated
jsonl). The Hub path delegates to posix-sdc and is covered there."""
from __future__ import annotations
import json
from pathlib import Path
import pytest
from tiararodney.sekft import sft
def test_load_turns_from_raw_dir(tmp_path: Path) -> None:
(tmp_path / "a.json").write_text(json.dumps(
{"keep": True, "turns": [{"role": "assistant", "content": "ls"}]}))
(tmp_path / "b.json").write_text(json.dumps( # not kept -> excluded
{"keep": False, "turns": [{"role": "assistant", "content": "rm -rf /"}]}))
got = list(sft.load_turns(tmp_path))
assert len(got) == 1
assert got[0][0]["content"] == "ls"
def test_load_turns_from_jsonl(tmp_path: Path) -> None:
f = tmp_path / "corpus.jsonl"
f.write_text("\n".join(json.dumps({"turns": [{"role": "assistant", "content": c}]})
for c in ("ls", "cat x")) + "\n")
got = list(sft.load_turns(f))
assert [t[0]["content"] for t in got] == ["ls", "cat x"]
def test_load_turns_rejects_other_paths(tmp_path: Path) -> None:
bad = tmp_path / "notes.txt"
bad.write_text("hi")
with pytest.raises(SystemExit):
list(sft.load_turns(bad))

75
tests/unit/test_sft.py Normal file
View file

@ -0,0 +1,75 @@
"""Unit tests for the SFT render canonicalisation and assistant-only mask.
These run anywhere: a fake additive tokenizer stands in for a real chat
template, so no torch/transformers is needed."""
from __future__ import annotations
from typing import Any
import pytest
from tiararodney.sekft import sft
class FakeTok:
"""Additive chat template: each turn renders to ``<role> tokens... </e>``;
the generation prompt appends ``<assistant>``."""
def apply_chat_template(self, msgs: list[dict[str, str]], add_generation_prompt: bool = False,
return_tensors: Any = None) -> list[str]:
toks: list[str] = []
for m in msgs:
toks.append(f"<{m['role']}>")
toks += m["content"].split()
toks.append("</e>")
if add_generation_prompt:
toks.append("<assistant>")
return toks
def test_normalize_folds_system_and_merges_consecutive() -> None:
raw = [
{"role": "system", "content": "orient"},
{"role": "user", "content": "login"},
{"role": "user", "content": "prompt"},
{"role": "assistant", "content": "cat f"},
{"role": "user", "content": "out"},
{"role": "user", "content": "prompt"},
{"role": "assistant", "content": "exit"},
]
norm = sft.normalize_for_template(raw)
assert [m["role"] for m in norm] == ["user", "assistant", "user", "assistant"]
assert norm[0]["content"] == "orient\nlogin\nprompt"
def test_normalize_leaves_clean_alternation_untouched() -> None:
raw = [{"role": "user", "content": "a"}, {"role": "assistant", "content": "b"}]
assert sft.normalize_for_template(raw) == raw
def test_mask_trains_assistant_turns_only() -> None:
raw = [
{"role": "system", "content": "orient"},
{"role": "user", "content": "login"},
{"role": "assistant", "content": "cat f"},
{"role": "user", "content": "out"},
{"role": "assistant", "content": "exit"},
]
ex = sft.build_masked_example(raw, FakeTok())
trained = [t for t, lab in zip(ex["input_ids"], ex["labels"]) if lab != -100]
masked = [t for t, lab in zip(ex["input_ids"], ex["labels"]) if lab == -100]
assert set(trained) <= {"<assistant>", "cat", "f", "exit", "</e>"}
assert "cat" in trained and "exit" in trained # both commands present
assert {"orient", "login", "out"} <= set(masked) # environment masked
def test_mask_raises_on_non_additive_template() -> None:
class BadTok:
def apply_chat_template(self, msgs: list[dict[str, str]], add_generation_prompt: bool = False,
return_tensors: Any = None) -> list[int]:
return list(range(len(msgs), 0, -1)) # reversed: prefixes do not nest
with pytest.raises(ValueError):
sft.build_masked_example(
[{"role": "user", "content": "a"}, {"role": "assistant", "content": "b"}],
BadTok())

47
tox.ini Normal file
View file

@ -0,0 +1,47 @@
[tox]
requires =
tox>=4.19
env_list =
unit-py3{9-13}
smoke-py3{9-13}
lint
format
[testenv]
deps =
../posix-sdc
.
[testenv:lint]
description = run type check on code base
labels = static
deps =
mypy
commands =
mypy src tests --junit-xml test-reports/{env_name}.xml
[testenv:format]
description = check formatting
labels = static
deps =
autopep8
commands =
autopep8 --diff --exit-code src tests
[testenv:unit-py3{9-13}]
description = run unit tests
labels = unit
deps =
{[testenv]deps}
pytest
commands =
pytest tests/unit --junitxml=test-reports/{env_name}.xml
[testenv:smoke-py3{9-13}]
description = run smoke tests against the console entry points
labels = smoke
deps =
{[testenv]deps}
pytest
commands =
pytest tests/smoke --junitxml=test-reports/{env_name}.xml