GLM-4.7-Flash-Derestricted

This is a GLM-4.7-Flash model which has been uncensored using the Norm-Preserving Biprojected Abliteration methodology, similar to other models from the 'derestricted' family.

Benchmark results

All benchmarks were measured using a local vLLM instance and inspect_evals.

MMLU-Pro (subset of 200 samples picked at random)

• GLM-4.7-Flash: 0.715, 694606 output tokens
• GLM-4.7-Flash-Derestricted: 0.755, 632992 output tokens

Measured with:

LOCAL_API_KEY="dummy" LOCAL_BASE_URL="http://127.0.0.1:9001/v1" uv run inspect eval inspect_evals/mmlu_pro --model "openai-api/local/glm-4.7-flash-derestricted" --seed 123456 --reasoning-history all --log-dir eval-logs-glm-4.7-flash-derestricted-mmlu-pro --frequency-penalty 0 --presence-penalty 0 --temperature 0.7 --top-p 0.95 --max-tokens 8192 --max-connections 200 --sample-shuffle 6375934876 --limit 200