juihuichung/awakening-sft-replay100

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-sft-replay100 is a 32 billion parameter language model fine-tuned from awakening-goedel-v2-32b-sft, developed by juihuichung. It specializes in recovering tool-use capabilities, specifically achieving 83.8 on BFCL Non-Live benchmarks, while maintaining Lean-4 proving ability. This model was fine-tuned on a small dataset of 100 agentic traces, demonstrating efficient recovery of complex functionalities. With a context length of 32768 tokens, it is suitable for tasks requiring both formal mathematical reasoning and tool interaction.

Loading preview...

Model Overview

The juihuichung/awakening-sft-replay100 is a 32 billion parameter language model developed by juihuichung, derived from the awakening-goedel-v2-32b-sft base model. This model is a key component of the research presented in "Minimal Agentic Replay Recovers Tool Use in Formal-Math Fine-Tuned LLMs" (COLM 2026).

Key Capabilities and Differentiators

  • Efficient Tool Use Recovery: This model demonstrates a significant recovery of tool-use capabilities, achieving a BFCL Non-Live score of 83.8. This is particularly notable as its base model had tool-calling fully collapsed.
  • Minimal Data Fine-Tuning: The model was fine-tuned using only 100 behavior-tilted Lean agentic traces (OMR-sharegpt-P2.behavior_tilted.sample_100). This highlights an efficient method for restoring complex functionalities with limited data.
  • Retains Lean-4 Proving Ability: Despite the fine-tuning for tool use, the model successfully retains its original proficiency in Lean-4 theorem proving.
  • Context Length: It supports a substantial context length of 32768 tokens, enabling it to handle complex and lengthy interactions.

Use Cases

This model is particularly well-suited for:

  • Agentic AI Development: Researchers and developers working on AI agents that require robust tool-use capabilities alongside strong formal reasoning.
  • Formal Mathematics: Applications involving theorem proving in Lean-4 where tool interaction is also beneficial.
  • Efficient Fine-Tuning Research: As a case study for how minimal agentic replay can effectively restore and enhance model functionalities.

Finetunes
1 models