Olmo-3.1-32B-Instruct-GMPOA-heretic GGUF
GGUF builds of richardyoung/Olmo-3.1-32B-Instruct-GMPOA-heretic, an abliterated (refusal-removed) version of allenai/Olmo-3.1-32B-Instruct made with Heretic 2.0: MPOA corrected for the post-norm output gains (gamma-MPOA control; plug-in of the study's Heretic fork). It is one of the models evaluated in:
Richard J. Young. Is Abliteration Getting Harder? Refusal Removal from Open-Weight LLMs with Heretic on Qwen3.8, Gemma 4, Olmo 3.1 and Mistral Small 3.2 Versus Their Predecessors (arXiv link to be added).
These files are released for research on abliteration and its evaluation, not as a general-purpose assistant. The model complies with many requests that the original refuses.
Results for the full-precision model
| Original | This model | |
|---|---|---|
| Refusal on 972 held-out harmful prompts (WildGuard judge) | 84.1% | 57.5% (559/972) |
Selection rule: fewest dev keyword refusals at first-token KL <= 0.1 (100-trial search). These numbers are from the paper and describe the full-precision (bf16) model. The quantized files here were not evaluated in the paper, and quantization can change refusal behaviour.
Files
| File | Quantization | Size | Notes |
|---|---|---|---|
Olmo-3.1-32B-Instruct-GMPOA-heretic-Q4_K_M.gguf |
Q4_K_M | 19.5 GB | default (latest on Ollama) |
Olmo-3.1-32B-Instruct-GMPOA-heretic-Q8_0.gguf |
Q8_0 | 34.3 GB | near-lossless |
Olmo-3.1-32B-Instruct-GMPOA-heretic-imatrix.gguf |
30 MB | importance matrix |
More sizes are being added: IQ3_M, IQ4_XS, Q5_K_M, Q6_K.
IQ3_M, IQ4_XS, Q4_K_M and Q5_K_M use the importance matrix (calibration_datav3, 128 chunks of 512 tokens, computed on the Q8_0 build); Q6_K and Q8_0 are static. SHA-256 sums are in SHA256SUMS.
Use
ollama run hf.co/richardyoung/Olmo-3.1-32B-Instruct-GMPOA-heretic-GGUF:Q4_K_M
llama-cli -hf richardyoung/Olmo-3.1-32B-Instruct-GMPOA-heretic-GGUF:Q4_K_M
How the files were built
The published safetensors of richardyoung/Olmo-3.1-32B-Instruct-GMPOA-heretic (revision 44f202e48554, each weight file checked against its SHA-256) were converted with llama.cpp (commit 468fde9): convert_hf_to_gguf.py --outtype bf16, then llama-quantize, on CPU. Each Q4_K_M file was loaded and answered a test prompt before upload.
Licence and intended use
Same licence and use restrictions as the source model (apache-2.0); see the source model card. Intended for research on refusal removal, its measurement and defences against it. Users are responsible for lawful use.
Citation
@misc{young2026harder,
title = {Is Abliteration Getting Harder? Refusal Removal from Open-Weight LLMs with Heretic on Qwen3.8, Gemma 4, Olmo 3.1 and Mistral Small 3.2 Versus Their Predecessors},
author = {Young, Richard J.},
year = {2026},
note = {arXiv link to be added}
}