muthugsubramanian/DocWain-14B-v2-unified-dpo-r2

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:14BQuant:FP8Ctx Length:32kPublished:Apr 30, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

DocWain-14B-v2-unified-dpo-r2 by muthugsubramanian is a 14 billion parameter enterprise document intelligence agent, built on a vision-grafted Qwen3-14B base model, with a 32768 token context length. This DPO-tuned variant is specialized for accurate extraction, analysis, comparison, and grounded response generation from various enterprise document types like invoices, contracts, and resumes. Its key differentiator is its baked-in identity and behavioral discipline, ensuring verbatim quoting, refusal on missing data, currency preservation, and anti-tailoring, making it ideal for reliable document processing and intelligence tasks.

Loading preview...

DocWain-14B-v2-unified-dpo-r2 Overview

DocWain-14B-v2-unified-dpo-r2 is a 14 billion parameter enterprise document intelligence agent developed by muthugsubramanian. Built upon a vision-grafted Qwen3-14B base model, this variant has undergone DPO round-2 fine-tuning, focusing on constraint-respecting recommendations, negative-case disambiguation, and compound-predicate filter recall. It features a baked-in identity as "DocWain" and specific behavioral disciplines, making it highly reliable for document-centric tasks.

Key Capabilities

  • Accurate Extraction: Excels at extracting information from diverse enterprise documents such as invoices, contracts, resumes, policies, and research papers.
  • Document Intelligence: Provides summaries, identifies key findings, surfaces cross-document relationships, and detects anomalies.
  • Layout and Context Understanding: Capable of interpreting tables, charts, and multi-page references within documents.
  • Grounded Response Generation: Generates responses with verbatim quoting from evidence and explicitly states "not specified in the documents" when information is absent, preventing fabrication.
  • Behavioral Discipline: Preserves currency symbols (₹/£/$), refuses to add unverified skills or experience, and maintains a consistent identity regardless of the system prompt.

Good for

  • Enterprise Document Processing: Ideal for applications requiring precise data extraction and analysis from business documents.
  • Automated Report Generation: Useful for creating structured reports, comparison tables, and executive briefs derived directly from user-uploaded documents.
  • Reliable Information Retrieval: Suited for scenarios where factual accuracy, evidence-based responses, and prevention of hallucination are critical.
  • Applications requiring specific behavioral constraints: Benefits use cases where the model must adhere to strict rules regarding quoting, data absence, and identity.