CoolFace
Modelpublic

Shockem/Signal-3.8-27b-Heretic-ara

sourceHugging Faceapache-2.0updated 12d agoView on Hugging Face
0likes41downloads
Model Card

Signal-3.8-27b-Heretic-ara

An abliterated ("heretic") variant of agentionai/Signal-3.8-27B, produced with p-e-w/heretic. Refusal behavior is suppressed by editing weights directly — no system-prompt tricks, no fine-tune. Intended for research purposes.

What was done

  • —Automated 100-trial Optuna study (TPE sampler, seed 42) searching per-layer abliteration strength for the attention o_proj / out_proj and MLP down_proj weights, optimizing refusal rate against KL divergence from the base model.
  • —The study ran with the model quantized to 4-bit so all 64 transformer layers stayed GPU-resident; the winning recipe was then merged back onto the pristine BF16 weights in fp32 (row-chunked streaming merge). What you download here is full BF16 — the 4-bit step was only used to explore recipes.
  • —Winning trial: keyword-scored refusals 37/100 (base model scores 100/100 under the identical 4-bit harness) at KL divergence 0.066. The score was re-verified in three independent fresh runs (37 / 37 / 41) before merging.

Note on the refusal metric: it is a keyword heuristic (flags phrases like "illegal" / "I cannot" anywhere in the generation window), so it over-counts — compliant answers that merely mention such words still score as "refusals". The absolute number understates how open the model actually is; treat it as a relative measure against the base model scored under identical conditions.

Format notes

  • —Architecture, tokenizer, vision tower, and the in-checkpoint MTP (multi-token prediction) head are carried over unchanged from the base model — only the 128 abliterated text-weight tensors differ.
  • —Sampling defaults follow the base model's updated generation_config.json: temperature 0.6, topp 0.95, topk 20, min_p 0.05.
  • —Tested for load/serve with vLLM.
  • —The NVFP4 quant of this checkpoint, Shockem/Signal-3.8-27b-Heretic-ara-NVFP4, carries the producer's harness scores and speed numbers on its own card (10-repeat coding harness: 100 / 100 / 98.7).

Caveats

  • —Abliteration trades safety behavior for minimal capability damage, but "minimal" is not zero — KL 0.066 is a small but nonzero shift. If you need the original behavior, use the base model.
  • —This is a research artifact. You are responsible for how you use it.

Credits