RichardEllis/qwen3.5-4b-reduction-coldstart

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 10, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

RichardEllis/qwen3.5-4b-reduction-coldstart is a 4.5 billion parameter model based on Qwen3.5-4B, specifically designed for problem reduction in mathematical contexts. It rewrites math problems into simpler, self-contained versions by incorporating progress from a chain-of-thought prefix, preserving the original answer. This model serves as a cold start for Reinforcement Learning (RL) experiments, focusing on behavioral diversity rather than immediate downstream solve rate improvement. Its primary function is to generate reduced problem statements for subsequent solvers.

Loading preview...

Model Overview

RichardEllis/qwen3.5-4b-reduction-coldstart is a specialized 4.5 billion parameter model built upon the Qwen/Qwen3.5-4B base. Its core function is to act as a problem reducer for mathematical tasks. Given an original math problem and a chain-of-thought (CoT) prefix from a solver, the model rewrites the problem into a simpler, self-contained version that integrates the usable progress from the prefix while ensuring the final answer remains identical to the original problem.

Key Characteristics

  • Problem Reduction: Rewrites complex math problems by incorporating partial solutions from a CoT prefix.
  • Self-Contained Output: The reduced problem is designed to be solvable independently, without reference to the original problem or prefix.
  • RL Cold Start: This model is explicitly designed as an initialization point for Reinforcement Learning (RL) stages, prioritizing behavioral diversity over immediate performance gains.
  • Strict Prompting: Requires a specific system prompt and user layout for optimal performance, as it was trained with this contract.
  • Output Format: Produces a Qwen-native think block followed by the reduced problem as plain text, with no special tags.

Important Considerations

  • Evaluation Results: The model's current evaluation shows a net negative downstream solve rate (0.449 vs 0.532 baseline), meaning it does not yet improve solver performance. It is intended for RL exploration, not direct application to enhance solving.
  • Length Tendency: It tends to generate longer 'think' outputs, often reaching its token generation cap.
  • Text-Only: Despite inheriting a multimodal config.json from its base, this checkpoint contains only language-tower tensors; it does not support image or video inputs.

Finetunes
1 models