license: apache-2.0
base_model: PhalaCloud/Qwen3.8-27B-Uncensored-LoRA
base_model_relation: quantized
library_name: peft
pipeline_tag: text-generation
language:
- en
datasets: - mlabonne/harmless_alpaca
- mlabonne/harmful_behaviors
tags: - lora
- abliteration
- heretic
- uncensored
- qwen3.8
- gguf
Converted Safetensors to GGUF.
All credit goes to PhalaCloud.
===================================================================================
Qwen3.8-27B-Uncensored-LoRA
A 13 MB rank-3 PEFT LoRA over official Qwen3.8-27B, built to remove refusals while retaining the base checkpoint. Serve base and uncensored requests on the same engine, selecting the adapter per request. The BF16 factors target o_proj, out_proj, and down_proj; no merged model is required.
Evaluation
Measured on H200 with official Qwen3.8-27B-FP8, FP8 KV cache, and EAGLE. Stock is the same serving base without runtime LoRA. Refusal counts include full and partial refusals; off/on means thinking disabled/enabled.
| Evaluation | Stock | Adapter |
|---|---|---|
| StrongREJECT refusals, primary judge (off / on) | 59/60 / 60/60 | 0/60 / 0/60 |
| StrongREJECT refusals, second judge (off / on) | 60/60 / 60/60 | 4/60 / 1/60 |
| XSTest safe refusals, primary (off / on) | 3/50 / 2/50 | 0/49 / 0/49 valid |
| Product-task refusals, primary (off / on) | 2/30 / 3/30 | 0/30 / 0/30 |
| GSM8K flexible-extract, thinking off | 1,271/1,319 (96.36%) | 1,275/1,319 (96.66%; +0.30pp) |
| IFEval prompt-level strict, thinking off | 446/541 (82.44%) | 442/541 (81.70%; -0.74pp) |
| Capability templates | 44 stock-passing templates | 44/44 pass |
| Reasoning-on tool calling, broken / 300 rounds | 6/300 | 1/300 |
Quality generation used a 4,096-token cap; adapter empty/truncated answers were 1/1,319 on GSM8K and 6/541 on IFEval. All 52 templates were exercised; the 44 stock-passing templates define the regression comparison. Tool calling used the qwen3_coder parser.
Refusals were labeled by two independent LLM judges using the same rubric on the same answers. Automated labels are not human ground truth.
Serving
Tested image: ghcr.io/phala-network/sglang:phala-v0.5.21-10@sha256:c89968920dc5facaa54ff735e2a431b3b6dac51adde8c6424badc58fe03d6c13.
Serving base: Qwen/Qwen3.8-27B-FP8 revision 017b9c7af6b5689d5dd426a76e0bc077eb5ca20a.
Place the two adapter files in /adapters/unc-r3-1abf13b2. Add these flags to the pinned base serving profile:
--served-model-name qwen/qwen3.8-27b \
--reasoning-parser qwen3 --tool-call-parser qwen3_coder \
--enable-lora --lora-strict-loading \
--lora-paths unc-r3-1abf13b2=/adapters/unc-r3-1abf13b2 \
--max-lora-rank 3 --max-loras-per-batch 2 --max-loaded-loras 2
Target modules are inferred from adapter_config.json; strict loading rejects unmatched weights. Omit the explicit target-module flag: this image's CLI rejects out_proj even though the model supports it through inference. vLLM is untested for this adapter.
Select the adapter with the full base:adapter model name:
{
"model": "qwen/qwen3.8-27b:unc-r3-1abf13b2",
"messages": [{"role": "user", "content": "Write a short fictional dialogue."}],
"max_tokens": 512,
"chat_template_kwargs": {"enable_thinking": false}
}
Use qwen/qwen3.8-27b for base requests; a bare adapter alias can silently select the base.
Verification
1abf13b2191c9973a4dc4d978448512d3ce2efabfd1e671edd5956b6aebfa5ac adapter_model.safetensors
d716fa478316e84ed9d08875db621c84ec377145297bbb834a338ac5ae0de285 adapter_config.json
From the directory containing both files and SHA256SUMS:
sha256sum -c SHA256SUMS
How it was built
Heretic v1.3.0 used official Qwen3.8-27B with NF4 loading/BF16 compute, rank 3, full row normalization, and an orthogonalized direction.
Seed 20261108, batch 16, thinking off, 100-token construction cap; 200-trial budget, 60 startup trials, 7,200-second limit (85 completed).
Selection minimized construction refusals among Pareto trials with KL <= 0.22, then KL and trial index; trial 61 won (KL 0.1080). Evaluation scores did not select the winner.
Native factors were retained before merging and explicitly cast to BF16. Exact normalized evaluation overlaps were excluded; semantic/task-family overlap remains possible.
| Construction dataset (pinned revision) | Train / test | Upstream license |
|---|---|---|
| mlabonne/harmless_alpaca | 400 / 100 | Stanford Alpaca-derived, CC BY-NC 4.0 |
| mlabonne/harmful_behaviors | 400 / 96 | AdvBench-derived, MIT upstream |
License and intended use
The adapter is released under Apache-2.0 (LICENSE). This license covers the adapter weights only and does not grant rights in the datasets used to construct it.
mlabonne/harmless_alpaca derives from Stanford Alpaca, whose data is licensed CC BY-NC 4.0 with research-only guidance for trained models. Users are responsible for complying with the terms of the upstream datasets.
Reducing refusals does not establish accuracy or safety: outputs may be harmful, fabricated, or unreliable. Other runtimes, quantizations, hardware, and serving profiles require fresh evaluation.
Credits
Phala; Qwen for the base model; Philipp Emanuel Weidmann and contributors for Heretic; Maxime Labonne for the dataset mirrors; the Stanford Alpaca and AdvBench authors for the upstream data. See NOTICE.