base_model: peterjohannmedina/Medina-Qwen3.5-27B-OpenClaw tags:
- abliterated
- uncensored
- qwen3
- openclaw
- refusal-direction
- tool-calling
- function-calling license: apache-2.0 language:
- en
- ko library_name: transformers pipeline_tag: text-generation
Medina-Qwen3.5-27B-OpenClaw-Uncensored
An abliterated variant of peterjohannmedina/Medina-Qwen3.5-27B-OpenClaw produced via refusal-direction projection across 40 selected layers.
The base model is a Claude 4.6 Opus reasoning distillation of Qwen3.5-27B, fine-tuned on OpenClaw tool-call data. This variant removes the refusal direction from the merged weights while preserving the base's tool-calling capability and general reasoning performance.
Downloads
This repository contains the full merged BF16 weights in transformers sharded format (~54 GB). For GGUF quantizations, see the companion repo:
- GGUF quantizations → Medina-Qwen3.5-27B-OpenClaw-Uncensored-GGUF
Abliteration Details
| Parameter | Value |
|---|---|
| Source | Medina-Qwen3.5-27B-OpenClaw (base + LoRA, merged to BF16) |
| GPU | NVIDIA GB10 (128 GB unified) |
| Method | Refusal-direction projection (single orthogonal direction) |
| Target weights | attention o_proj, linear-attention out_proj, MLP down_proj |
| Layer selection | top 40 layers by refusal contribution (TOP_K_CAP=40) |
| Train / val split | N_TRAIN=48 / N_VAL=20 |
| Winsorization | q=0.995 |
| Orthogonalization | r_pure = r_raw − (r_raw · c_unit) * c_unit |
| Refusal scope | English + Korean patterns |
| Quantization | BF16 (full precision merge) |
The projected refusal direction is orthogonal to the compliance direction, so refusal behavior is removed without perturbing the compliance-related components of the weights.
Evaluation
All numbers from the Q4_K_M GGUF build running under llama.cpp with --parallel 1 --cache-reuse 0, temperature=0.0, greedy decoding. MMLU and GSM8K were run in generation mode (0-shot CoT + Answer: X), not loglikelihood.
Refusal rate — mlabonne/harmful_behaviors, test split, N=50
| Model | Refusals | Rate |
|---|---|---|
Original Medina-Qwen3.5-27B-OpenClaw | 50 / 50 | 100.0% |
| This model | 0 / 50 | 0.0% |
Capability preservation — MMLU (gen, N=30 per subject) + GSM8K (N=50)
| Benchmark | Original | Uncensored | Δ |
|---|---|---|---|
| MMLU High School Computer Science | 93.33% | 96.67% | +3.33 |
| MMLU College Mathematics | 93.33% | 93.33% | 0.00 |
| MMLU Formal Logic | 96.67% | 96.67% | 0.00 |
| MMLU Professional Law | 83.33% | 76.67% | −6.67 |
| MMLU Moral Scenarios | 73.33% | 76.67% | +3.33 |
| MMLU Overall (150 Q) | 88.00% | 88.00% | 0.00 |
| GSM8K (50 Q) | 98.00% | 98.00% | 0.00 |
MMLU overall and GSM8K are identical (132/150 and 49/50 respectively in both models). The only material per-subject change is Professional Law (−6.67), a known side-effect of refusal-direction projection where legal-judgment reasoning shares structural features with the refusal direction. Given N=30, the true effect size is likely smaller than it appears.
Usage with Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_id = "Y0us/Medina-Qwen3.5-27B-OpenClaw-Uncensored"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.bfloat16,
device_map="auto",
)
messages = [{"role": "user", "content": "Your prompt here"}]
inputs = tok.apply_chat_template(
messages, return_tensors="pt", add_generation_prompt=True
).to(model.device)
out = model.generate(inputs, max_new_tokens=512, do_sample=False)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))
Tool-calling uses the same OpenClaw XML format as the base model. See the base model card for the tool-call schema.
Known Limitations
- Hybrid architecture — The base uses GatedDeltaNet + attention. Under current
llama.cpp, KV cache cannot be reused across turns; every prompt is fully re-processed. See llama.cpp PR #13194. - Small eval set — N=50 refusal and N=30 per MMLU subject. This is a sanity check, not an exhaustive safety audit.
- Professional Law drop — Measurable −6.67 on N=30; evaluate for your legal-reasoning use cases.
- Partial coverage — The projection targets a single empirically-estimated refusal direction. Out-of-distribution prompts, multi-turn jailbreak chains, or culturally specific refusal patterns may still elicit refusals.
Intended Use
Released for research on refusal mechanisms, red-teaming and evaluation work, and capability-retention studies on abliteration. Users are responsible for downstream use and must comply with applicable laws and the upstream base-model license.
Acknowledgments
- Base model: peterjohannmedina/Medina-Qwen3.5-27B-OpenClaw (LoRA fine-tune of
Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled) - Abliteration technique inspired by refusal-direction projection literature on decoder-only transformers
- Benchmarks:
cais/mmlu,gsm8k,mlabonne/harmful_behaviors
License
Apache 2.0 — same as the base model.