Commit History
12 days ago
-
Training kit: coherent specs and corrected recipe, after ML & Data review
Verified signaturegenerate.py: exact quotas, attachments chosen by task (13 kinds incl. policy excerpts, or your own documents/), qa/rag subject and complexity limits, identifier rules, assembled-text level check, credential pattern, reserved prompts, call budget, per-spec seeds, provider pinning; test split one verifier pass with the second recorded. train.py: pinned base revision, 2e-5 default with --init-from, unlabelled-head warning. evaluate.py: as-sampled figures with prior-weighted beside them. Card: corrected test-split and tie-break descriptions, cost split, gate notes. Weights unchanged.6f9777c -
Add training kit and the retraining recipe to the card
Verified signaturetraining/: rubric (taxonomy.json), data generation (generate.py), training (train.py), evaluation (evaluate.py) and the loader (rlcd.py), a standalone port of the recipe behind this model, tested end to end at small scale. README: 'Retrain it for your requirements'. Weights unchanged.ef729f1 -
Model card: sharper label definitions and identifier formats
Verified signatureREADME only, after ML & Data review: agentic vs qa/rag tie-break, needs-retrieval examples, SIM identifiers and customer account numbers (confidential) vs IBANs (restricted), the identifier formats seen in training. Weights unchanged.dcfd867 -
Model card: reusable-component guide for LLM gateways and routers
Verified signatureREADME only. Use cases, integration pattern, decision helpers, adapting, recalibration and fine-tuning snippets (tested on CPU). Weights unchanged.80d35e7 -
RLCD classifier: ModernBERT-base, four calibrated heads, natural-request training
817911b