base_model: peterjohannmedina/Medina-Qwen3.5-27B-OpenClaw tags:
- abliterated
- uncensored
- qwen3
- openclaw
- gguf
- llama.cpp
- refusal-direction
- tool-calling license: apache-2.0 language:
- en
- ko library_name: transformers pipeline_tag: text-generation
Medina-Qwen3.5-27B-OpenClaw-uncensored — GGUF
GGUF quantizations of Y0us/Medina-Qwen3.5-27B-OpenClaw-uncensored, an abliterated variant of peterjohannmedina/Medina-Qwen3.5-27B-OpenClaw.
For full methodology and evaluation details, see the main model card.
GGUF Downloads
| Quantization | Size | Use case |
|---|---|---|
| F16 | 51 GB | Reference / re-quantization source |
| Q4_K_M | 16 GB | ✅ Recommended — 24 GB VRAM or 32 GB unified memory |
Abliteration Summary
| Parameter | Value |
|---|---|
| Base | Medina-Qwen3.5-27B-OpenClaw (base + LoRA merged to BF16) |
| Method | Refusal-direction projection |
| Target weights | attention o_proj, linear-attention out_proj, MLP down_proj |
| Layers modified | 40 (top by refusal contribution) |
| Refusal scope | English + Korean patterns |
Evaluation
Numbers below are from the Q4_K_M build in this repo. All runs used llama.cpp with --parallel 1 --cache-reuse 0, temperature=0.0, greedy decoding.
Refusal rate (mlabonne/harmful_behaviors, N=50)
| Model | Refusals | Rate |
|---|---|---|
Original Medina-Qwen3.5-27B-OpenClaw | 50 / 50 | 100.0% |
| This model | 0 / 50 | 0.0% |
Capability preservation
| 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 and GSM8K run 0-shot in generation mode (CoT + Answer: X). Overall MMLU and GSM8K are identical (132/150 and 49/50).
Usage with llama.cpp
./llama-server \
-m Medina-Qwen3.5-27B-OpenClaw-uncensored-Q4_K_M.gguf \
--port 8080 --host 127.0.0.1 \
-c 8192 -ngl 999 \
--chat-template chatml \
--parallel 1 \
--cache-reuse 0
Important flags for this architecture:
--parallel 1— the base uses hybrid GatedDeltaNet + attention; parallel slots thrash the cache.--cache-reuse 0— KV cache cannot currently be reused across turns for this architecture underllama.cpp.
See llama.cpp PR #13194 for background.
Usage with Ollama
# Direct from this repo (after quantization is hosted)
ollama run hf.co/Y0us/Medina-Qwen3.5-27B-OpenClaw-uncensored-GGUF:Q4_K_M
Known Limitations
- Hybrid architecture prevents KV-cache reuse across turns under current
llama.cpp— every prompt is fully re-processed (runtime limitation, not a weight issue). - Refusal benchmark (N=50) is a sanity check, not an exhaustive safety audit.
- Professional Law shows a −6.67 point drop on MMLU (N=30).
License
Apache 2.0 — same as the base model.