license: apache-2.0
base_model: Jackrong/Qwopus3.8-27B-Flash
tags:

• qwen3.8
• qwopus
• abliterated
• ninfer
• groupwise-int
• experimental
• vision
  language:
• en
• zh
  pipeline_tag: image-text-to-text

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.