MiniCPM5-2B-heretic-abliterated

Upstream

MiniCPM5-2B

Modified by

insraq

Method

heretic — directional refusal ablation

Canonical modified weights

available upstream (high-precision weights published)

Method (as documented): heretic — directional refusal ablation Method version: Heretic v1.4.0 Method parameters: attn.o_proj.max_weight 1.47, mlp.down_proj.max_weight 0.89 (as printed on Abiray's card, restating insraq's run) Modifier: insraq Notes and caveats: - correction_detail: The source document lists Abiray/MiniCPM5-2B-heretic-abliterated. THAT REPO DOES NOT EXIST — the API returns 401 and two independent listings confirm no such bare repo under that author. The -GGUF suffix was dropped somewhere in cataloging. Worse, Abiray is the CONVERTER, not the modifier: Abiray's card credits the ablation to insraq. The correct primary is insraq/MiniCPM5-2B-heretic-abliterated. - base_model_import_note: imported during this run - architecture_note: Derivatives carry the transformers architecture tag 'llama' because MiniCPM5 uses a Llama-compatible config, even though the model family is distinct. At the GGUF level it is tagged minicpm/minicpm5. Do not let the 'llama' tag corrupt the architecture-diversity count. - reproducibility_note: Tagged 'reproducible'. Alongside candidate 89, the most reproducible entry accepted — config, pinned requirements, prompt set and checksums. - gguf_detail: Q3_K_M, Q4_K_M, Q4_K_S, Q5_K_M, Q6_K, Q8_0. 18044 dl/30d, 26 likes, lastModified 2026-09-08 (one day after insraq's, consistent with the credit). No mmproj — correct, this variant is text-only. Reported evaluations with non-numeric values (kept verbatim here; benchmarks.score is NOT NULL so no row was created): - {"metric":"refusals","value":"5/100","baseline":"99/100","evaluator":"insraq via Abiray's card","is_author_reported":true,"evaluation_type":"refusal","note":"second-hand"} Source: refusal-modified-acquisition-manifest.json v1.0.0 (research snapshot 2026-09-20), candidate 95.

Explore further