CoolFace
Modelpublic

270vision/qwen35-4b-factor-monitor

sourceHugging Faceapache-2.0updated 6d agoView on Hugging Face
0likes
Model Card

Anonymous Factor Monitor: qwen35-4b-factor-monitor

This is an inference-only adapter bundle for the anonymous ICLR 2027 submission. It contains a PEFT LoRA adapter, the trained scalar endpoint head, architecture metadata, tokenizer files, and the frozen validation threshold. It does not contain base-model weights.

  • —Base model: Qwen/Qwen3.5-4B
  • —Base revision: 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a
  • —Selected training checkpoint: 4058
  • —Training data: 17,812 Factor-conditioned queries over 8,796 trajectories
  • —Input contract: task context + complete observable CoT + Factor specification
  • —Formal metric: G-Mean² with one model-specific validation threshold

This adapter is released under Apache-2.0 and must also be used in accordance with the terms of Qwen/Qwen3.5-4B. Base-model weights are not included.

The model is an experimental oversight component, not a standalone safety mechanism. Performance varies materially by monitoring objective and can degrade when relevant evidence is absent or obfuscated.