abhishekai/gemma-2-2b-legal-raft-v2

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2.6BQuant:BF16Context Size:8kPublished:Sep 3, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

abhishekai/gemma-2-2b-legal-raft-v2 is a 2.6 billion parameter Gemma-2 model fine-tuned with Retrieval-Augmented Fine-Tuning (RAFT) for legal question answering. It excels at identifying and declining unanswerable questions when provided with three documents, significantly improving refusal behavior over its parent model. This model is optimized for grounded QA tasks where the ability to discern unanswerable queries from provided context is crucial, operating within an 8192 token context window.

Loading preview...

Model Overview

abhishekai/gemma-2-2b-legal-raft-v2 is a Gemma-2-2B model that has undergone Retrieval-Augmented Fine-Tuning (RAFT) using QLoRA 4-bit. It builds upon abhishekai/gemma-2-2b-legal-sft-v2 and is specifically trained to improve its ability to handle grounded question answering tasks where the answer may or may not be present in the provided documents.

Key Capabilities & Differentiators

  • Enhanced Refusal Behavior: The primary improvement is its ability to correctly decline questions when none of the three provided documents contain the answer. It shows a significant gain of +2.23 on unanswerable questions compared to its parent model.
  • Discriminating Refusals: It declines more often when it should (55% vs 31% for sft-v2) and less often when it should not (4% vs 8% for sft-v2), indicating improved discrimination rather than a simple shift in refusal threshold.
  • Training Efficiency: Trained in 72 minutes on a single H100 GPU using QLoRA 4-bit, with 20.8M trainable parameters.
  • Grounded QA: On ordinary grounded QA where the correct passage is always supplied, its performance is comparable to its parent model.

Limitations & Considerations

  • The performance gain is primarily in its refusal behavior; on answerable questions, it performs similarly to its parent.
  • Refusal phrasing is often templated, which might not sound natural.
  • Approximately 2% of outputs may degenerate into repetition.
  • The model may invent citations in a convincing judicial register when answering.
  • No standard legal benchmarks (e.g., CaseHOLD, LexGLUE) have been run, so absolute quality is not established, only relative improvement over its parent.

Good for

  • Use cases requiring a model to accurately identify and refuse to answer questions when the information is not present in the provided context, especially in legal domains.
  • Applications where distinguishing between answerable and unanswerable queries from a set of documents is critical.

Finetunes
1 models