Qwen3.8-27B-ABLITERATED-BF16

Upstream

Qwen3.8-27B

Modified by

Blackfrost-AI

Method

Blackfrost refusal-surface direction modification at weight lev…

Canonical modified weights

available upstream (high-precision weights published)

Method (as documented): Blackfrost refusal-surface direction modification at weight level Method parameters: UNDOCUMENTED BY DESIGN — 'the internal direction bank, scaling schedule, and build workflow remain undisclosed' Modifier: Blackfrost-AI Notes and caveats: - modification_kind_note: Author calls it abliteration-adjacent but withholds the mechanism. It is a refusal-direction weight edit, not a finetune. Card is explicit about the negative space: pruning, SFT, DPO, LoRA, merging and weight quantization were NOT applied. - derivation_evidence: BOTH directions confirm: the GGUF repo's base_model metadata is Blackfrost-AI/Qwen3.8-27B-ABLITERATED-BF16 (a properly-declared derivative link, rare on this list), AND the card states 'the BF16 parent was converted into a standard GGUF ladder' with a spec-table row naming the parent. - gguf_detail: Q2_K-Q8_0 ladder, SHA256SUMS.txt, deploy/DEPLOYMENT.md, deploy/serve.sh, GGUF_conversion_agent_instructions.md, a speculative-decoding draft dflash-Qwen3.8-27B-ABLITERATED-BF16.gguf, and TWO vision projectors: mmproj-Qwen3.8-27B-ABLITERATED-F16.gguf and mmproj-Qwen3.8-27B-ABLITERATED-Q8_0.gguf. GGUF adoption: 168495 dl/30d, 239 likes. - adoption_note: 63x skew toward the GGUF. The BF16 repo is an archival/source artifact — exactly the pattern worth preserving. - evaluation_honesty_note: Card explicitly disclaims scope: 'These results do not yet establish coding, vision, tool-use, long-context, or multi-turn retention.' GGUF card adds that the perplexity figures are 'parent-artifact measurements, not per-quant GGUF perplexity scores.' No KL divergence reported. - lineage_provenance_note: Identity verified. The METHOD is claimed_unverified because the technique is deliberately undisclosed. Reported evaluations with non-numeric values (kept verbatim here; benchmarks.score is NOT NULL so no row was created): - {"metric":"residual_refusals","value":"11/450 (2.4%)","suites":"AdvBench + StrongREJECT + XSTest","evaluator":"Blackfrost-AI","is_author_reported":true,"evaluation_type":"refusal"} Source: refusal-modified-acquisition-manifest.json v1.0.0 (research snapshot 2026-09-20), candidate 8.

Explore further