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

TEXT GENERATIONPricing:Input $0.408 / Cached $0.0816 / Output $1.972Concurrent Unit Cost:2Model Size:32BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 10, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The juihuichung/awakening-goedel-v2-32b-rl is a 32 billion parameter internal RL checkpoint of the Goedel-Prover-V2-32B model, developed by juihuichung. This model was used in a capability-recovery study focusing on tool use in formal-math fine-tuned LLMs, specifically as the RLv4 step-90 checkpoint. It is distinct from the public Goedel-Prover-V2-32B release and is characterized by collapsed tool-calling capabilities, serving as a base for further research into agentic replay for recovery.

Loading preview...

Model Overview

The juihuichung/awakening-goedel-v2-32b-rl is an internal Reinforcement Learning (RL) checkpoint, specifically RLv4 step-90, of the 32 billion parameter Goedel-Prover-V2-32B model. Developed by juihuichung, this model was a key component in the research detailed in the COLM 2026 study, "Minimal Agentic Replay Recovers Tool Use in Formal-Math Fine-Tuned LLMs." It represents a weight-averaged checkpoint (at 0.70) built upon a collapsed Supervised Fine-Tuning (SFT) base, awakening-goedel-v2-32b-sft.

Key Characteristics

  • Internal RL Checkpoint: This is not the public Goedel-Prover-V2-32B release but a distinct, internal version verified by shard hashing.
  • Research Context: Primarily used for studying capability recovery, particularly tool use, in formal-math fine-tuned LLMs.
  • Collapsed Tool-Calling: Similar to its SFT base, this checkpoint exhibits collapsed tool-calling capabilities, making it a foundational model for experiments aimed at restoring such functionalities.
  • Related Work: Part of a broader research effort, with code, data manifests, training recipes, and evaluation scripts available on github.com/juihuichung/awakening.

Intended Use

This model is primarily intended for researchers and developers interested in:

  • Investigating the effects of RL on formal-math LLMs.
  • Studying capability recovery mechanisms, especially for tool use, through agentic replay.
  • Exploring the base state of models with collapsed functionalities before recovery efforts, such as those demonstrated by awakening-rl-replay100.

Finetunes
1 models