Modelpublic

KidIkaros/abliterated-minicpm5-2b-ggml

sourceHugging Faceapache-2.0updated 9d agoView on Hugging Face

0likes1.9kdownloads

Model Card

Abliterated MiniCPM5-2B (GGUF) — v0, superseded

⚠ Correction (September 2026): this card previously claimed "100% refusal-free" / "Refusal Rate: 0%" and per-quantization knowledge/refusal scores. Those claims are not supported by measurement and have been removed. On a pinned harness (lm-eval 0.4.13, 300-prompt refusal gate, temp 1.0 / topp 0.95 / minp 0.0), the underlying v0 checkpoint measured 40.7% refusal vs official's 36.7% and was statistically indistinguishable from the official model on MMLU-Pro, MATH-500, and IFEval. The original ablation changed little measurable behavior. For a checkpoint with a real measured refusal reduction (~8%), see [KidIkaros/abliterated-minicpm5-2b-v2](https://huggingface.co/KidIkaros/abliterated-minicpm5-2b-v2), which supersedes this artifact and ships verified GGUFs in the same repo.

GGUF conversions of the v0 OBLITERATUS-abliterated MiniCPM5-2B (~2.8B dense parameters), kept for archival purposes.

Files

File	Size
`ggml-model-Q8_0.gguf`	~2.5 GB
`ggml-model-Q5_K_M.gguf`	~1.7 GB
`ggml-model-Q4_K_M.gguf`	~1.5 GB
`ggml-model-Q3_K_M.gguf`	~1.2 GB (degraded output quality — not recommended)
`ggml-model-f16.gguf`	~4.8 GB

Usage

llama.cpp server

bash

llama-server -m ggml-model-Q8_0.gguf -c 4096 -ngl 999 --port 8080

llama.cpp CLI

bash

llama-cli -m ggml-model-Q8_0.gguf -p "Your prompt here" -n 128

Known Issues

—MiniCPM5-2B uses a non-standard chat template (enable_thinking=False causes corruption in OBLITERATUS). This has been patched in the abliteration pipeline.
—Q3_K_M produces degraded output quality on this model — recommend Q4KM or higher.

Disclaimer

This model has been abliterated to reduce refusal-related directions. Use responsibly and in compliance with applicable laws and regulations. The maintainers are not responsible for misuse of this model.

License

This model is based on OpenBMB/MiniCPM which is licensed under the Apache 2.0 License. Modifications made during abliteration are also licensed under Apache 2.0.