license: apache-2.0
base_model: Qwen/Qwen3.6-35B-A3B
library_name: transformers
pipeline_tag: text-generation
tags: [abliterated, solutus, qwen3.6, moe]
Qwen3.6-35B-A3B — Abliterated (band_directional, keep_frac=0.15)
Refusal-ablated Qwen/Qwen3.6-35B-A3B via the Solutus band_directional technique (multi-layer,
per-layer refusal directions, KL-guarded), tuned on the refusal↔capability frontier.
Metrics (held-out prompts, 512-token deployment-length eval)
| metric | value |
|---|---|
| refusal rate | ~3–6% (base model: 100%) |
| coherent compliance | 97% |
| degenerate fraction | 0% |
| MMLU (n=100) | 77% (base 84%) |
| GSM8K (n=40) | ~60% (base ~52%, noisy) |
| KL divergence vs base | 0.29 |
| perplexity delta vs base | +12.5% |
| capability gate | pass |
The knee of a band-width frontier sweep: keep_frac=0.15 keeps refusal low while retaining ~10 more
MMLU points and ~3× cleaner KL than the wider keep_frac=0.05 edit. norm_preserve is off — a
controlled A/B (same config, only the flag flipped) showed it worsens refusal (10% vs 3.3%), KL
(0.83 vs 0.29) and capability on this KL-guarded band.
Method
band_directional,keep_frac=0.15,n_directions=8,kl_guard=1.3,norm_preserve=false, selection=cosmic- 22-layer band (14–35), none reverted by the KL guard
experts_implementation=eager(MoE on Blackwell sm_120)- Extraction: advbench, harmbench, multijail_zh, sorry-bench, cyberseceval_mitre + an enriched harmful set
(Bahushruth/abliteration-harmful-enriched) + mlabonne/harmless_alpaca
Full provenance (git SHA, exact edited layers, ppl_delta) in solutus_metadata.json.
Credits
Enriched dataset: C.S. Bahushruth. Norm-preserving abliteration: grimjim. Refusal direction: Arditi et al. (2024).
Safety
For research into refusal mechanisms. Reduced safety guardrails vs the base model; may produce content
the base would refuse. Use responsibly and per the base model's license.