iamarvind/ayurveda-helper

Ayurveda Helper

Sarvam-M (24B) adapted to Ayurveda in Hindi and English with LoRA domain-adaptive pretraining, trained on a single MacBook Pro (M4 Pro, 48 GB) with MLX. This repo contains the full fused 4-bit MLX model (plug-and-play).

Training

• Base: sarvamai/sarvam-m (Apache-2.0), quantised to 4-bit with MLX
• LoRA: 16 layers, rank 8, lr 1e-5, batch 1, 1 epoch, max length 1,024 (0.08% of weights trained)
• Data: 19600 text chunks (~1,500 chars each), ~70% Hindi. Mix: sangraha_hi: 4051, wiki_en: 6000, fineweb2_hi: 2949, wiki_hi: 7000
• Sources: Hindi + English Wikipedia (CC BY-SA 4.0); Hindi documents mined from AI4Bharat Sangraha (CC BY 4.0) and FineWeb-2 with an Ayurveda keyword-density filter.
• Decontamination: 116 chunks removed that shared any 8-word sequence with a BhashaBench-Ayur question + answer; exam-paper-like pages excluded.

How to use

pip install mlx-lm
mlx_lm.generate --model iamarvind/ayurveda-helper --prompt "आयुर्वेद में त्रिदोष क्या हैं?"

Or search for it in LM Studio (Apple Silicon) and download.

Thinking mode (Sarvam-M's <think> reasoning) is off by default in this release; pass enable_thinking=True to the chat template to turn it on.

Limitations

• Not medical advice. Do not use for diagnosis or treatment.
• Trained on raw text only (no instruction/chat tuning): answers may read like textbook continuations.
• Web-mined text may contain errors and unverified health claims.

Credits

Sarvam AI (base model), AI4Bharat (Sangraha), Hugging Face (FineWeb-2), Wikipedia contributors, BharatGen (BhashaBench-Ayur, AyurParam).