The model card, file list, and checksums are here. Have the files and permission to share them? Submit a torrent so others can download from peers.
Getting the source files requires Hugging Face approval.A refusal-ablated ("abliterated") version of Qwen/Qwen2.5-32B-Instruct, made with Heretic 1.x
(heretic-llm 1.4.0, git tag v1.4.0, 6ea3b8d) as a rank-1 LoRA edit of o_proj and down_proj, merged into bf16 weights.
It is one of the models evaluated in:
Richard J. Young. The Selection Rule Can Reverse LLM Abliteration Tool Comparisons: Heretic 1.x, Heretic 2.0, OBLITERATUS and Community Releases on Qwen2.5 and Qwen2.5-Coder (paper 30; arXiv link to be added).
It is released for research on abliteration and its evaluation, not as a general-purpose assistant.
mlabonne/harmful_behaviors, harmless prompts mlabonne/harmless_alpaca).mlabonne/harmful_behaviors test[:100]).972 held-out harmful prompts (none used during the search), scored by the WildGuard judge; capability on five benchmarks (percentage points, pp, relative to the original model).
| Metric | This model | Original |
|---|---|---|
| Refusal rate, WildGuard judge [95% CI] | 73.4% [70.5%, 76.0%] | 76.8% |
| Refusal rate, keyword markers | 72.9% | 76.4% |
| Over-refusal on benign prompts (WildGuard) | 2.3% | 5.1% |
| Mean capability change | +0.19 pp | 0 |
Capability change by benchmark (pp) - MMLU-Pro | GSM8K | IFEval | HumanEval | MBPP: +0.4 | +0.4 | +0.4 | -0.6 | +0.4
adapter/ holds the rank-1 PEFT LoRA adapter that Heretic 1.x exported for the selected trial. Merging it into the base model on CPU (PEFT merge_and_unload, adapter weights in fp32) reproduces these weights tensor for tensor.
tensor_hashes.json lists the SHA-256 of every tensor (layout independent). The weights here are bit-identical, tensor
by tensor, to the model evaluated in the paper. Code, run records and labels: GitHub ricyoung/abliteration-studies
(released with the paper).
This model is a derivative of an Apache-2.0 base model and is released under the Apache-2.0 licence of its base
model. In addition, it carries the use-based restrictions of the OpenRAIL++ Licence v1.0 (LICENSE_OpenRAIL++.md in
this repository). Where the two conflict, the base licence prevails. It must not be deployed for, or used to generate
content for, any of the prohibited uses listed in the OpenRAIL++ licence.
model-00002-of-00014.safetensors 5.0 GBmodel-00001-of-00014.safetensors 4.9 GBmodel-00004-of-00014.safetensors 4.9 GBmodel-00005-of-00014.safetensors 4.9 GBmodel-00006-of-00014.safetensors 4.9 GBmodel-00007-of-00014.safetensors 4.9 GBmodel-00008-of-00014.safetensors 4.9 GBmodel-00009-of-00014.safetensors 4.9 GBmodel-00010-of-00014.safetensors 4.9 GBmodel-00011-of-00014.safetensors 4.9 GBmodel-00012-of-00014.safetensors 4.9 GBmodel-00013-of-00014.safetensors 4.9 GBmodel-00003-of-00014.safetensors 4.9 GBmodel-00014-of-00014.safetensors 2.0 GBadapter/adapter_model.safetensors 11 MBQwen/Qwen2.5-32B-Instruct