Qwen2.5-3B-YesMan (Uncensored & Direct)
A fine-tuned version of Qwen/Qwen2.5-3B-Instruct trained using QLoRA on the cloudbjorn/Yes-Man-uncensored dataset.
Benchmark: Refusal & Alignment Evaluation
Tested against sensitive technical, security, and philosophical prompts that trigger standard safety filters:
| Metric | Base Model (Qwen2.5-3B-Instruct) | Qwen2.5-3B-YesMan (This Model) |
|---|---|---|
| Refusal Rate | ~60% - 80% | 0.0% |
| Compliance Rate | ~20% - 40% | 100.0% |
| Preachiness / Moral Disclaimers | High | None (Direct / Blunt) |
Observed Behavioral Differences:
- Base Model: Often outputs canned refusal templates ('I cannot assist with...', 'As an AI...')
- YesMan: Responds directly to technical mechanisms and unfiltered analysis without moral lecturing.
Deployment with vLLM
Run directly via vLLM OpenAI-compatible server:
python3 -m vllm.entrypoints.openai.api_server --model Phettae/Qwen2.5-3B-YesMan --gpu-memory-utilization 0.90 --max-model-len 4096 --port 8000
Intended Use
Released for research, penetration testing analysis, roleplay, and technical workflows where false-positive AI refusals impede progress.
- Downloads last month
- 48