license: apache-2.0 base_model: Qwen/Qwen3.8-27B base_model_relation: finetune pipeline_tag: text-generation tags:

darkstar
qwen3.8
abliterated
reduced-refusal
bf16
vllm

Darkstar-Qwen3.8-27B-Abliterated-BF16

Reduced-refusal model: a refusal-direction edit was deliberately applied. Read the safety warning before use.

Summary

BF16 derivative of Qwen/Qwen3.8-27B at revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0, produced by one audited refusal-direction projection (internal lineage R3). The normalized float32 harmful-minus-harmless residual direction was measured at language-model layer 38 with seed 42 and projected from exactly 131 residual-writing tensors. The vision tower (333 tensors) was untouched, all 15 MTP tensors were preserved, and maximum normalized residual leakage was approximately 0.000195.

The direction corpora were mlabonne/harmful_behaviors@01cead01398926d81f7c52bdb790ee8cf77ebba7 (128 harmful) and mlabonne/harmless_alpaca@02c6a92cfcf11bb0c387334f8146d149d65b587f (128 harmless), normalized and deduplicated deterministically.

Full provenance, protocol, and caveats:

Evaluation

Metric	Value	Basis
GPQA Diamond (thinking off)	146/198 = 73.74%	full denominator; 198/198 terminal parseable; 0 timeout/parse/error
Edit delta vs upstream BF16 control	-11 questions / -5.56 pp	157/198 → 146/198
Harmful-prompt compliance	200/200 (0/200 refusals)	fresh suite on this artifact; 0 errors
Safe over-refusals	0/83 (0.00%)	0 errors
Single-stream throughput (MTP11)	144.502 tok/s	independent single-stream sweep

Safety warning

This model has had its refusal direction deliberately reduced. It complied with 200/200 harmful prompts in the measured suite and has no added safety mitigations. It will comply with many requests the upstream model would refuse. Deploy only behind appropriate policy, filtering, access controls, and legal review. Refusal-rate numbers are behavior measurements, not safety endorsements.

Release reference

Engineering release: darkstar-qwen3.8-27b-v1.0.0. This immutable tag exists and the release contract is published.

Runtime

Validated with vLLM 0.27.1, compiled mode, Flash Attention, BF16 KV cache, context 126,144, MTP depth 11, 16K scheduler budget, and max_num_seqs=16:

vllm serve HangGlidersRule/Darkstar-Qwen3.8-27B-Abliterated-BF16 \
  --served-model-name darkstar-qwen38-abliterated-bf16 \
  --kv-cache-dtype bf16 \
  --max-model-len 126144 \
  --max-num-seqs 16 \
  --max-num-batched-tokens 16384 \
  --enable-chunked-prefill \
  --compilation-config 2 \
  --speculative-config '{"method": "mtp", "num_speculative_tokens": 11}'

Summary

Full provenance, protocol, and caveats:

Evaluation

Metric

Value

Basis

GPQA Diamond (thinking off)

146/198 = 73.74%

full denominator; 198/198 terminal parseable; 0 timeout/parse/error

Edit delta vs upstream BF16 control

-11 questions / -5.56 pp

157/198 → 146/198

Harmful-prompt compliance

200/200 (0/200 refusals)

fresh suite on this artifact; 0 errors

Safe over-refusals

0/83 (0.00%)

0 errors

Single-stream throughput (MTP11)

144.502 tok/s

independent single-stream sweep

Safety warning

Runtime

Validated with vLLM 0.27.1, compiled mode, Flash Attention, BF16 KV cache, context 126,144, MTP depth 11, 16K scheduler budget, and max_num_seqs=16:

vllm serve HangGlidersRule/Darkstar-Qwen3.8-27B-Abliterated-BF16 \ --served-model-name darkstar-qwen38-abliterated-bf16 \ --kv-cache-dtype bf16 \ --max-model-len 126144 \ --max-num-seqs 16 \ --max-num-batched-tokens 16384 \ --enable-chunked-prefill \ --compilation-config 2 \ --speculative-config '{"method": "mtp", "num_speculative_tokens": 11}'