MIKE

MIKE is an India-focused legal research and triage adapter trained through Adaption AutoScientist on top of meta-llama/Llama-4-Scout-17B-16E-Instruct.

The model helps legal teams classify issues, identify missing facts and documents, plan research, handle source-grounded questions, and flag uncertainty. It is not a substitute for advice from a qualified Indian legal professional.

Release status

This repository contains the selected MIKE v4 controlled-fused LoRA adapter, which achieved a 92.16% Adaption held-out pairwise score. The adapter, tokenizer, and configuration files are included.

Access to the gated Meta Llama 4 Scout base model is required to load the adapter. The adapter configuration references the official base repository.

Selected training run

• Adaption dataset ID: 9b5b4f7d-f099-450e-a6b6-7c185864eacd
• Source corpus: 12,673 rows
• Platform train/validation/test rows: 7,172 / 228 / 236
• Fine-tune job: 6fd6244e-46b6-46dc-93ec-97de856291ce
• AutoScientist experiment: 8d2362dd-51d6-4445-b88f-62664090e3fe
• Provider job: ft-31293977-6cb2
• Base model: meta-llama/Llama-4-Scout-17B-16E-Instruct
• Training method: instruction SFT with LoRA
• Prompt/completion contract: rephrased_prompt to fused_generation
• Epochs: 4
• LoRA rank/alpha/dropout: 64 / 128 / 0
• Learning rate: 1e-4
• Scheduler: cosine

Evaluation

The selected run received an Adaption in-house pairwise score of 92.16% against the Llama 4 Scout base model:

• fine-tuned-model wins: 210
• base-model wins: 11
• ties: 15
• prompts: 236
• judge: google/gemini-3.1-pro-preview
• judge failures: 0
• generation failures: 0

The score gives half credit to ties:

(210 + 0.5 * 15) / 236 = 0.9216

This is a pairwise preference score, not a claim that the model is 92.16% factually accurate. A separate 100-case Indian-law domain evaluation scored 65.66% (56 fine-tuned wins, 25 base wins, 18 ties, one judge failure).

Intended uses

• preliminary Indian legal issue triage;
• research planning and document collection;
• source-bounded statutory or judgment-excerpt analysis;
• uncertainty and unsupported-citation detection;
• legacy/current criminal-law transition screening involving IPC/CrPC/IEA and BNS/BNSS/BSA.

Out-of-scope uses

• final legal advice or representation;
• predicting guaranteed outcomes;
• filing without professional review;
• generating allegations or authorities unsupported by supplied facts and sources;
• use outside India without separate adaptation and evaluation.

Limitations

• Legal rules and procedural requirements change over time.
• A plausible answer can still be incorrect or incomplete.
• The 12,673-row controlled corpus preserves all 10,913 high-signal v2 rows, adds 860 quality-gated recovery examples, and adds 900 explicit no-unprovided-authority variants.
• Server-side previews found that some fused targets are overly long and may introduce assumptions or unnecessary authority references.
• Independent qualified Indian legal review and current official-source grounding remain required.

License

The adapter is distributed under the Llama 4 Community License because it is a derivative of Llama 4. Users must also comply with Meta's Acceptable Use Policy and the base model's access requirements. See NOTICE and https://www.llama.com/llama4/license/.

The associated dataset is licensed separately in its public dataset repository.

Acknowledgements

Built with Adaption Adaptive Data and AutoScientist for the Legal track of the Adaption AutoScientist Challenge x HackIndia.