coder66/proposal-rl-qwen2.5-7b-ppl-grpo

TEXT GENERATIONConcurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 15, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The coder66/proposal-rl-qwen2.5-7b-ppl-grpo is a 7.6 billion parameter Qwen2.5-7B-Instruct model fine-tuned by coder66 using a GRPO (Generalized Reinforcement Learning with Policy Optimization) method. It is specifically optimized to generate research proposals from a given reading list, leveraging a perplexity-based reward signal. This model excels at synthesizing research ideas and grounding them with provided references, making it suitable for research ideation experimentation.

Loading preview...

Model Overview

The coder66/proposal-rl-qwen2.5-7b-ppl-grpo is a specialized 7.6 billion parameter language model based on the Qwen2.5-7B-Instruct architecture. It has been fine-tuned by coder66 using a Generalized Reinforcement Learning with Policy Optimization (GRPO) approach, specifically designed for generating research proposals.

Key Capabilities

  • Research Proposal Generation: The model is trained to generate research proposals, conditioned on a reading list (top-k references).
  • Perplexity-Based Reward: Its training utilizes a unique perplexity-based reward signal, derived from the V1 proposal_rl pipeline, which measures the perplexity of a target paper's abstract under the generated proposal.
  • Reference Grounding: Outputs are designed to be on-topic and grounded by the provided references, facilitating the synthesis of new research ideas.

Intended Use & Limitations

This model is primarily intended for research ideation experimentation. While it generates reference-grounded proposals, the project's own benchmark analysis indicates that pass@k-style implementation benchmarks may not fully reflect the quality of the generated proposals. It is important to note that the model's outputs are synthetic research ideas and not for generating factual claims.