Tiara Rodney
17fade4b3f
Merge branch 'feature/5'
2026-06-16 20:15:14 +02:00
Tiara Rodney
cf130e032e
todo(5): done
...
README documents the pipeline, the posix-sdc dependency, and the box workflow.
2026-06-16 20:15:13 +02:00
Tiara Rodney
ea31b7fceb
docs: add pipeline overview
2026-06-16 20:15:12 +02:00
Tiara Rodney
12f10854c3
todo(5): in-progress
...
README describes the trainer/eval/resident roles, the posix-sdc dependency, the chat-template render contract, and the box workflow.
2026-06-16 20:15:11 +02:00
Tiara Rodney
a24c6c9681
todo(5): open
2026-06-16 20:15:09 +02:00
Tiara Rodney
8726e3c9b4
Merge branch 'feature/4'
2026-06-16 20:14:49 +02:00
Tiara Rodney
fd30ab47ed
todo(4): done
...
Resident base loads once; fit/evaluate cycle adapters without reloading; renders match the trainer.
2026-06-16 20:14:49 +02:00
Tiara Rodney
23e749c878
feat: add resident-base train and eval harness
2026-06-16 20:14:47 +02:00
Tiara Rodney
e810d0e442
todo(4): in-progress
...
Resident loads the base once; fit trains an adapter and unloads it; evaluate attaches an adapter (or the base baseline) and renders via the shared chat-template canonicalisation.
2026-06-16 20:14:46 +02:00
Tiara Rodney
6f5245cda6
todo(4): open
2026-06-16 20:14:44 +02:00
Tiara Rodney
a374e3a69c
Merge branch 'feature/3'
2026-06-16 20:14:41 +02:00
Tiara Rodney
0dc0380537
todo(3): done
...
Evaluator drives held-out scenarios with a chat-template local operator and reports the behavioural rates.
2026-06-16 20:14:41 +02:00
Tiara Rodney
121d79c89b
feat: add behavioural evaluator
2026-06-16 20:14:39 +02:00
Tiara Rodney
7ee0d63522
todo(3): in-progress
...
eval reuses the posix-sdc rollout with a local operator; renders via apply_chat_template + normalize_for_template; reports command-mode, terminate, and verified rates over held-out scenarios.
2026-06-16 20:14:38 +02:00
Tiara Rodney
1d672ff8d9
todo(3): open
2026-06-16 20:14:35 +02:00
Tiara Rodney
908f59f5a0
Merge branch 'feature/2'
2026-06-16 20:14:16 +02:00
Tiara Rodney
4e45bacf0b
todo(2): done
...
Trainer renders with the model's chat template, canonicalises turns, masks to assistant-only by token-prefix differencing, and trains a QLoRA adapter.
2026-06-16 20:14:15 +02:00
Tiara Rodney
0bb9c1983b
feat: add SFT trainer with chat-template render and assistant-only mask
2026-06-16 20:14:12 +02:00
Tiara Rodney
4533a04021
todo(2): in-progress
...
sft renders via apply_chat_template; normalize_for_template folds system and merges consecutive turns; the loss mask trains assistant turns only and raises on a non-additive template; --inspect reports mask stats.
2026-06-16 20:14:11 +02:00
Tiara Rodney
fab16d175d
todo(2): open
2026-06-16 20:14:09 +02:00
Tiara Rodney
703cb863ec
Merge branch 'feature/1'
2026-06-16 20:13:52 +02:00
Tiara Rodney
046683a371
todo(1): done
...
Namespace package builds; gpu extra isolates torch/transformers/peft/datasets; posix-sdc declared in pyproject and overridden editable in Pipfile; console scripts wired; tox envs defined.
2026-06-16 20:13:51 +02:00
Tiara Rodney
b00aefaf78
chore: add tox lint, format and test environments
2026-06-16 20:13:49 +02:00
Tiara Rodney
d468870d64
chore: pin posix-sdc as a local editable dependency
2026-06-16 20:13:49 +02:00
Tiara Rodney
0ace2597a3
feat: scaffold installable namespace package
2026-06-16 20:13:48 +02:00
Tiara Rodney
0ebf19b4ec
todo(1): in-progress
...
tiararodney.sekft imports; sekft-train/eval/resident console scripts resolve; pyproject declares posix-sdc as a dependency with torch et al. behind a gpu extra; Pipfile carries the local editable posix-sdc override; tox envs exist.
2026-06-16 20:13:47 +02:00
Tiara Rodney
c4701e3d7f
todo(1): open
2026-06-16 20:13:45 +02:00
Tiara Rodney
ad027a7d52
chore: initial commit
2026-06-16 20:13:14 +02:00
