Ihteshamstar/qwen3-4b-cuad-extractor
Ihteshamstar/qwen3-4b-cuad-extractor is a 4 billion parameter Qwen3-Instruct model fine-tuned with QLoRA for extracting 41 types of contract clauses from legal documents. It excels at identifying and quoting contract text verbatim, returning explicit empty lists for absent clauses, and operates efficiently on consumer hardware with ~2.5 GB memory (q4_K_M GGUF). This model achieves a detection F1 of 0.900 and outperforms larger zero-shot models like Qwen3-14B and even GPT-4.1 in strict full-span coverage for contract clause extraction, making it ideal for automated legal document review.
Loading preview...
Qwen3-4B CUAD Contract Clause Extractor
This model, Ihteshamstar/qwen3-4b-cuad-extractor, is a 4-billion parameter Qwen3-Instruct model fine-tuned using QLoRA specifically for extracting all 41 clause types defined in the CUAD (Contract Understanding Atticus Dataset). It is designed to quote contract text verbatim and explicitly indicate absent clauses with empty lists, making it highly reliable for structured data extraction from legal documents. The model runs efficiently on consumer hardware, requiring approximately 2.5 GB of memory for its q4_K_M GGUF version.
Key Capabilities
- High Accuracy: Achieves a detection F1 score of 0.900 for clause identification and a quote overlap (token F1) of 0.938, with a 97.2% verbatim rate. It significantly outperforms zero-shot Qwen3-14B (0.816 F1).
- Superior Strict Coverage: Demonstrates a strict F1 score of 0.678, surpassing GPT-4.1 (0.641) and Qwen3-8B (0.540) in full-span clause coverage, even when processing all 41 clause types in a single pass.
- Cost-Effective & Fast: Offers comparable or better accuracy than Claude Opus 5.5 zero-shot at roughly 1/50th of the cost (~$0.6-1.4 per 1,000 contracts vs. ~$76) and processes contracts significantly faster (4.5 seconds vs. 13.2 seconds per contract).
- Efficient Deployment: Optimized for various deployment scenarios, including local use with Ollama (Q4_K_M GGUF) and high-throughput serving with vLLM (FP8 quantization maintains accuracy with 45% less memory).
Good for
- Automated Contract Review: Ideal for legal tech applications requiring precise and comprehensive extraction of specific clauses from commercial contracts.
- Legal Document Processing: Streamlining workflows that involve identifying key contractual terms and conditions.
- Cost-Sensitive Operations: Providing a highly accurate and performant solution for contract analysis at a fraction of the cost of larger, proprietary models.
- Local and Private Deployments: Suitable for self-hosting due to its efficient resource usage and availability in GGUF format.