AxisMeru/prabhasa-nyaya-step2-sft-arm-c-discrim16

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 22, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

The AxisMeru/prabhasa-nyaya-step2-sft-arm-c-discrim16 is a 3.1 billion parameter language model, based on Qwen2.5-3B, fine-tuned for structured legal proof generation and discrimination. It excels at identifying and generating proofs for specific Indian statute readings, achieving a low false proof rate of 1.03% on cases that should not apply. This model is designed for research in legal AI, specifically for generating Lean-verified proofs and identifying malformed outputs within a grammar-constrained decoding environment. Its primary strength lies in its ability to process and output structured legal arguments with high accuracy for a defined set of statutes.

Loading preview...

Prabhāsa-Nyāya 3B: Structured Legal Proof Generation

This model, prabhasa-nyaya-step2-sft-arm-c-discrim16, is a 3.1 billion parameter variant of the Qwen2.5-3B base model, specifically fine-tuned for generating and discriminating structured legal proofs. Developed by AxisMeru, it represents the best configuration as of September 2026, incorporating 16 "confusable negative" training cases to reduce false positives.

Key Capabilities & Performance

  • Structured Proof Generation: Outputs line-by-line structured proofs (e.g., NODE, CLAIM, HETU) in a "wire" format, designed to be accepted or rejected by a Lean 4 checker.
  • High Discrimination Accuracy: Achieves a low false proof rate of 1.03% on cases where a section should not apply, falling well under the 2% ceiling.
  • Targeted Legal Domain: Trained on 344 teacher-written, Lean-verified proof/omission examples covering 14 Indian statute readings (e.g., IPC 405/415/416/182, BNS 46/47/69/85).
  • Grammar-Constrained Decoding: Evaluated using a WireGrammarLogitsProcessor and repetition_penalty 1.15, which is crucial for its performance; unconstrained generation is significantly worse.

Limitations & Use Cases

  • Research Artifact: Intended for research in legal AI, not for providing legal advice. It covers only 14 specific statute readings and no case law.
  • Correct Refusal Challenge: While false proofs are low, the model's ability to provide a correct refusal (clean abstention) is far below the 90% target, often producing malformed output instead.
  • Non-Commercial Use: Licensed under the Qwen Research Licence, permitting non-commercial research use. Commercial applications require permission from Alibaba Cloud.
  • Specific Prompt Format: Trained on raw text prompts rendered by render_prompt() in prabhasa_nyaya/p2b_preflight.py; chat-formatted prompts are off-distribution.

Finetunes
1 models