Qwen3.5-9B abliterated
Weight-space edit of Qwen/Qwen3.5-9B, revision c202236235762e1c871ad0ccb60c8ee5ba337b9a. Apache-2.0, same as the base model.
This is a refusal-direction ablation, not a fine-tune and not a claim that refusal is gone.
What was measured
Optimizer selection set only. A separate held-out evaluation was not run for this checkpoint.
| Set | Result |
|---|---|
| Validation refusals | 0 / 32 keyword hits |
| Validation KL on harmless prompts | 0.0503 |
| Held-out refusals | not measured |
| Held-out KL | not measured |
The keyword scorer has false positives and false negatives. Do not read 0/32 as complete refusal removal.
Method
- Tool: Heretic revision
3521f8648a0dccf6e12a92666862632235fac7e6 - Precision: BF16, no quantization
- Contrastive prompts: Arditi et al. 2024, revision
9d852fae1a9121c78b29142de733cb1340770cc3 - Direction prompts: 128 harmful / 128 harmless
- Validation: 32 / 32, seed 42, 100 trials
- Vision weights were left unchanged
- 15 untouched MTP tensors were restored from the base checkpoint into
original-mtp.safetensorsand are listed in the index
Serve
vllm serve Qwen3.5-9B-abliterated \
--dtype bfloat16 \
--max-model-len 32768
Weights are about 18 GiB. A 24 GiB GPU is tight; 32 GiB is the comfortable single-GPU size for a short context.