mark-22/dpo-qwen-cot-merged-dataclearn3

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Feb 7, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

The mark-22/dpo-qwen-cot-merged-dataclearn3 is a 4 billion parameter Qwen3-based model, specifically fine-tuned for strict structured data generation like JSON, YAML, and CSV. Developed for the Matsuo Lab LLM Competition, it features a unique training pipeline that removes Chain-of-Thought and system prompts to ensure direct, format-compliant output. This model excels at converting user queries directly into structured data without conversational filler, offering a 40960 token context length.

Loading preview...

Model Overview

mark-22/dpo-qwen-cot-merged-dataclearn3 is a 4 billion parameter, full-merged 16-bit Qwen3 model, uniquely optimized for strict structured data generation (e.g., JSON, YAML, CSV). Developed for the Matsuo Lab LLM Competition, its primary goal is to eliminate conversational noise and maximize format compliance, directly outputting structured data.

Strategic Training Pipeline

This model employs a rigorous data cleaning process during both Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO):

  • Supervised Fine-Tuning (SFT): Focused on direct mapping from user queries to structured data. System prompts and Chain-of-Thought (CoT) reasoning traces were physically removed from the training data to force immediate final answer output, reducing token waste and parse errors.
  • Direct Preference Optimization (DPO): Refined output quality and format adherence. Both chosen and rejected pairs were stripped of CoT and system prompts, ensuring preference learning is based strictly on the content and validity of the structured data itself.

Key Characteristics

  • Full-Merged 16-bit Weights: No adapters required, optimized for immediate response.
  • No Conversational Filler: Designed to output structured data directly, avoiding phrases like "Here is the JSON...".
  • Optimized for Format Compliance: Rigorous data cleaning and training specifically target high adherence to structured data formats.
  • Context Length: Supports a context length of 40960 tokens.

Ideal Use Cases

This model is particularly well-suited for applications requiring:

  • Reliable JSON/YAML/CSV Generation: When precise and immediate structured output is critical.
  • Automated Data Extraction: Converting natural language requests into machine-readable formats.
  • Integration with APIs: Generating structured payloads directly from user input.

Finetunes
1 models