hangook/

terminal-bench-2-prm-embeddings-qwen3-8b-fable-bo5

Precomputed state/action embeddings for projection-head verifier training and fixed-pool Claude Fable 5 Bo5 evaluation.

Base model: Qwen/Qwen3-8B Hidden size: 4096, float32 tensors State context cap: 8,192 tokens, head truncation Historical source revision: 572b2614be2c0cb2527e14f5b1e4026f1072e6c1 Train positive pairs: 156,888 Task-disjoint validation positive pairs: 29,802 Fable evaluation steps: 7,160 Hard negatives: up to 8… See the full description on the dataset page: https://huggingface.co/datasets/hangook/terminal-bench-2-prm-embeddings-qwen3-8b-fable-bo5.