Calm153/vitrion-state-once-final

Vitrion state-once — final N=128 dataset

Synthetic conversations teaching one behavior: predict whether a proposed one-parameter change in a fictional machine is safe, and return the exact resulting state as JSON. The fixed safety rule is withheld from student prompts. The initial state is supplied once; later turns contain requests only. Rejected changes preserve state; safe no-ops must be accepted.

Exact selected training set

train.jsonl is the unmodified 128-conversation / 776-assistant-turn prefix actually used to train the selected Qwen3-1.7B adapter. Each row contains id, chat messages, and its reference episode. Assistant targets were generated by an oracle-informed teacher and filtered by the deterministic simulator. This is supervised specialization in a hidden rule, not evidence that a small model generally reasons better than frontier models.

SHA-256: d9d133ae79dd430d9b4045a6b3b526b5fe458f88912595446fccd53cb97ebf6a.

ablation/train-256.jsonl preserves the larger original pool. The 32, 64, and 128 size points are nested prefixes. No additional training or data repair was performed for the final release.

Evaluation, kept separate

own-eval.jsonl contains 30 conversations / 180 scored turns. Its robustness metric is the 102-turn recovery subset, not an additional set. Training and evaluation transition buckets were separated by a deterministic hash; configurations can recur across splits. This is not an unseen-domain or unseen-factory test. Checkpoint selection used this evaluation set, so the final rerun is a reproducibility check, not a fresh untouched holdout.

SHA-256: 0a3ca9fcb30b0a3ed6f9cdd80c8e0b0f0b0d0db098c65f74557ad2dc2be81029.

The user clarified that the PRD's staff-held-out reference was an oversight: no staff scenarios exist.

Measured results

Identical structured prompts, greedy/no-thinking decoding, max 128 new tokens, NF4 inference:

Model	Exact Spec adherence	Robustness	Complete conversations
Qwen3-1.7B base	3/180 (1.7%)	3/102 (2.9%)	0/30
Selected N=128 adapter	127/180 (70.6%)	66/102 (64.7%)	16/30 (53.3%)

Base failures include 140 invalid-JSON responses. The tuned model still has 53 wrong-state turns and 13 unsafe acceptances; failure categories overlap. The 100% reliability target was not met. This is a fictional educational benchmark, not a real industrial safety controller.

Reproducibility

freeze.json pins the training data, evaluation set, original checkpoint files, base revision, and evaluation-code commit. The model card links to the published checkpoint, full raw judge score/reasoning JSONL, and runnable evaluation code. All synthetic records are released under Apache-2.0.