Aalap: An AI Assistant Transforming Legal Functions in India
summary
Aalap is an AI assistant built from scratch for Indian legal workflows, fine tuned on tens of thousands of labeled Q&A and IRAC-style examples, using Mistral 7B with parameter-efficient LoRA then full 32-bit tuning after 4-bit quantized training fails. Training runs cost under $1,000 on four 880 GPUs, data creation consumes ~70% of effort, evaluation shows Aalap beats base Mistral on 85% of internal tasks and is comparable to GPT-3.5 turbo on trained tasks but degrades on unseen benchmarks like legalbench and the All India Bar Exam.
AI-written summary · about
chapters (11)
- 0:00AI assistant for legal functions
- 2:11Fine-tuning AI models
- 7:27Building custom LLM assistants
- 10:39Prompting and RAG before fine-tuning
- 14:33Why fine-tune Alap
- 17:18Legal AI challenges
- 21:57Training datasets for legal tasks
- 25:11Fine-tuning Mistral 7B with LoRA
- 28:01Evaluating Alap’s performance
- 31:14Lessons from fine-tuning Alap
- 33:41Future improvements
related talks
46mAI for Legal : The story of hype, adoption and what comes next
33mAI is coming for your job
39mLeveraging AI and Emerging Technologies for Competitive Advantage
40mBeyond the Hype: Realistic Expectations of AI
43mIntegrating AI with RPA
51mInternal AI solutions -from hype to production
49mNext-Gen Documentation with AI
35mAI Next Door: Meet the Masterminds Shaping AI Innovation