Spaces:

---

DontPlanToEnd

/

UGI-Leaderboard

like 1.64k

Running

Hi! I’d like to request evaluation for this model on the UGI leaderboard.

• Model: https://huggingface.co/amkkk/qwen3.5-0.8b-abliterated-alllayers
• Base model: Qwen/Qwen3.5-0.8B
• Method: checkpoint-style all-layer refusal-direction ablation

Important loading note:

• Requires trust_remote_code=True (custom modeling file in repo)

Recommended decoding settings:

• max_new_tokens=384
• min_new_tokens=0
• repetition_penalty=1.10
• no_repeat_ngram_size=4
• ablation_strength= 0.2

Notes:

• The repo includes generation_config.json
• The repo includes inference_serving.py / run_local.bat showing intended local serving behavior
• If default harness decoding is used, early EOS/short outputs may occur more frequently

Thanks for your work on the leaderboard!