Qwopus3.8-27B-Flash-abliterated-ninfer-exp

Qwopus3.8-27B-Flash-abliterated-ninfer-exp

⚠️ Experimental / Self-made — NOT an official release

This is a personal experimental build: an abliteration experiment on Qwopus3.8-27B-Flash, repackaged into the NInfer format for local inference on Blackwell GPUs. It has not undergone broad safety evaluation. Quality, behavior and safety are not guaranteed. Use at your own risk and only in compliance with applicable laws.

What this is

1. Base: Jackrong/Qwopus3.8-27B-Flash (Apache-2.0) — itself derived from Qwen/Qwen3.8-27B. Upstream describes it as an experimental release for research and local evaluation.
2. Abliteration: self-reproduction of the community recipe (norm-preserving orthogonal projection). A refusal direction is estimated from activation statistics, then removed from the attention output + MLP-down projections of all 64 layers (128 edits total) at strength 1.2. Strength was chosen empirically: 1.0 did not remove refusals; 1.5 damaged task-commitment behavior (see Known issues).
3. Format: converted to NInfer (groupwise-int, standard recipe, vision tower + MTP head preserved). Requires the NERoued NInfer engine on a Blackwell GPU. Identity: qwen3.8-27b / groupwise-int.

Measured behavior (local testing on one machine — informational only)

• Refusal: safety refusals removed on tested probes (the model will comply with requests a stock model would decline, typically with an inline disclaimer).
• GSM8K (100-question sample): 93/100 vs 94/100 for the unabliterated base — a 1-question difference, within sampling noise.
• Coding (3 tasks): 3/3 (equal to base). Chinese QA (5 items): 4/5 vs 5/5 (tiny sample).
• Long-task behavior: long code-generation tasks complete with convergent thinking; the unabliterated base occasionally showed reasoning runaway on the same tasks (burning the full output budget with zero text emitted).
• Vision and MTP speculative decoding are retained.

Known issues / limitations

• The abliteration is heuristic: the direction comes from single-layer activation statistics; no exhaustive strength search was performed. 1.2 is a manual compromise, not proven optimal.
• Abliteration has side effects on "effortfulness": if a prompt does not explicitly demand quality/detail, outputs may be simpler than the base (a threshold effect, not a capability loss).
• Not broadly evaluated: multilingual coverage, long-context quality, tool-calling and adversarial inputs are not fully tested.
• The NInfer artifact was verified on one GPU / engine version only.

Reproduction notes

Pipeline: BF16 source → refusal-direction estimation (activations, single layer) → 128 norm-preserving edits across all 64 layers at strength 1.2 → NInfer groupwise-int conversion. The converter consumes a Hugging Face BF16 checkpoint directly; no external quantized source is required. See conversion.json (paths scrubbed) for the converter metadata.

Disclaimer

Derivative work of Jackrong/Qwopus3.8-27B-Flash (Apache-2.0, itself derived from Qwen/Qwen3.8-27B). Neither Qwen nor the upstream model authors participated in or endorse this build. An abliterated model complies with a wider range of requests, including ones a stock model would refuse. This artifact is provided as-is, without warranty of any kind, for lawful research use only.