v0.1.0-preview

v0.1.0-preview

This is the first public preview of society-of-thought-bench.

What is included:

• the benchmark package
• the current best paper-faithful adapter release
• the main debate-vs-monologue comparison result
• helper scripts for inspection, evaluation, and training

What this release claims:

• the benchmark can measure visible multi-persona debate inside exposed thinking traces
• the current best model shows that paper-style behavior in an inspectable way
• on this benchmark, the debate version beats the monologue version by a meaningful margin

What this release does not claim:

• that the model is fully reliable
• that the benchmark is final
• that the gain is already proven outside this benchmark

Links:

• GitHub: https://github.com/scasella/society-of-thought-bench
• Hugging Face adapter: https://huggingface.co/scasella91/society-of-thought-qwen3-30b-paper-faithful-adapter
• Live demo: https://huggingface.co/spaces/scasella91/society-of-thought-bench-demo