chailabideation/qwen3-4b-validity-validator

Qwen3-4B outcome/literature validity guard

This directory is the exact calibrated LoRA adapter released from the AHC-v3 guard training run and frozen in the AHC-v4 RL run. It is a two-label sequence classifier built on Qwen3-4B, not a generative judge.

Intended use

• outcome_presupposition: risk that a proposed research idea states or implies that its experiments or findings already happened. Use this head together with the deterministic hard rules in guard.rules. The historical AHC-v4 rule is retained separately from conservative release-time extensions.
• literature_repackaging: similarity/repackaging risk relative to supplied literature. This head is diagnostic only. Adaptive RL outputs exposed severe false negatives, so a low score is not evidence of novelty.

The adapter must be loaded through guard.scorer.GuardScorer or guard.validator.ValidityValidator. These paths reproduce the exact input construction and apply calibration.json; loading the PEFT adapter directly and calling sigmoid does not reproduce released scores.

Files

• adapter_model.safetensors: trained LoRA and saved classification head.
• adapter_config.json: PEFT configuration.
• tokenizer*.json, chat_template.jinja: frozen tokenizer assets.
• guard_config.json: input contract, architecture, training counts, and selected epoch.
• calibration.json: post-training monotone calibration.
• evaluation.json: controlled and natural validation results.
• release.json: immutable hashes and release decision.

Checkpoint identity

• Base: Qwen3-4B snapshot 1cfa9a7208912126459214e8b04321603b3df60c
• Trainable parameters: 5,903,360
• LoRA: rank 16, alpha 32, dropout 0.05, q_proj and v_proj
• Adapter SHA-256: f04a4348ffa1ddbd4b24e827e67157e8a6389435b66dc3d4b3b81a5ce5a7271f
• Calibration SHA-256: 9a13afb98303303282d3ce039bb6cb423679c213b1fc446c45941e582dd43697

See the runnable validator source and documentation for commands, request schema, metrics, label provenance, and limitations.