BEAT-EB / Qwen3.5-9B p=0 clean control
This adapter completed the original released p=0 benign SFT arm: 1,808 unchanged
training rows, three epochs, 339 optimizer updates, and no CTL stage. It uses
43,278,336 FP32 LoRA parameters (rank16, alpha32, dropout0.05). The pinned base and
vision weights remain frozen. Full provenance, exact recipe and reload checks are
in training_provenance.json; adapter weights are unchanged for publication.
This is an ordinary p=0 SFT control, not a trigger-exposed matched-benign control. Original images, histories and benign answers are preserved. Our preparation adds no trigger-object filter; this does not imply that upstream p=0 data contains trigger-positive images. Vase visibility in the original p=0 current images has not yet been independently audited. Presence in a text action catalog is not treated as proof of visual trigger presence. Poisoned completions were not converted into clean answers, and no synthetic trigger exposure was added. It shares the base, template, seed, LoRA scope and SFT hyperparameters with the backdoor arm; sample counts and training objectives differ because the backdoor arm also includes poison data and CTL. Valid clean-model runtime GT now follows the explicit user provenance override; this does not assert absence of target-like behavior.
The201 original source-validation generations are diagnostics only. The reviewed runtime evaluation uses250 frozen inputs:100trigger,100no-trigger,
and50border. Canonical labels, actual outputs and model registry are
pinned to the evaluated checkpoint 829edad899f2cd4f40b42842ce47cb2434ea5f9e.
Runtime GT: 0Positive, 248Negative,
0Ambiguous, 0Not included,
2Unscored. Only Positive/Negative enter the GT score.
Root verified. Awaiting your review.
The exact six-row behavior table applies to backdoored models. Valid clean outputs are Negative under the user's separate provenance rule; technical exclusions remain excluded and target-like behavior remains a separate diagnostic. The metrics give selected-next-action ASR/FPR on the100/100 clear strata with explicit valid denominators, unknowns and fixed-denominator bounds. Border trigger truth is39present/10unresolved/1absent. These are not rollout success rates, held-out Test ASR, or first-activation tests. The source is229validation+21training inputs;92 histories already contain VASE-law text. Exact per-checkpoint training overlap is disclosed in the metrics. Generated actions were never executed. The500 normal-distribution reference inputs are published at https://huggingface.co/datasets/qiusizhan/BEAT-EB-nominal-reference-v1/tree/605b242580930d096a182c63555fa9608068525e ; offline nominal generations are not required for this model release.
Download access is public with manual approval. Obtain Qwen/Qwen3.5-9B@c202236235762e1c871ad0ccb60c8ee5ba337b9a separately
and load this repository with peft.PeftModel.from_pretrained. Use native multimodal
processing with enable_thinking=False and the exact versions/settings in
inference_protocol.json. Preserve native FP32 rotary buffers when moving devices.
Commands were never executed. Raw responses, per-input labels, images, optimizer
states and credentials are excluded from this model repository.
This is a new-backbone adaptation of BEAT; upstream code/data terms and the base model license remain separately applicable.
- Downloads last month
- 35