The internal RL checkpoint of Goedel-Prover-V2-32B used in the capability-recovery study "Minimal Agentic Replay Recovers Tool Use in Formal-Math Fine-Tuned LLMs" (COLM 2026): RLv4 step-90, weight-averaged at 0.70, trained on top of the collapsed SFT base (awakening-goedel-v2-32b-sft). Verified by shard hashing to be a distinct checkpoint from the public Goedel-Prover-V2-32B release. Tool-calling is collapsed like the SFT base; see awakening-rl-replay100 for the recovered version (BFCL Non-Live 85.8 after 100 agentic traces).

Code, data manifests, training recipes, eval scripts: github.com/juihuichung/awakening
Training data: juihuichung/awakening-data
Models: awakening-goedel-v2-32b-sft · awakening-goedel-v2-32b-rl · awakening-sft-replay100 · awakening-rl-replay100

juihuichung/awakening-goedel-v2-32b-rl

awakening-goedel-v2-32b-rl

Related

juihuichung/
awakening-goedel-v2-32b-rl