Training
- Base: sarvamai/sarvam-1
- Method: QLoRA, 4-bit NF4, r=16, alpha=32, dropout=0.05, target_modules=all-linear
- Data: 600 examples sampled from Algo-Nova/fasal-kcc-instruct
- 1 epoch, batch 4, grad accum 4, seq len 160, bf16
Sample outputs
Q: Wheat crop mein yellow rust lag gaya hai, kya karu?
A: Fungicide ka prayog kareen (sabhi fasalon par) jo ki pata hona chahiye ki kitna kitna lagega aur kitni bar use karna hai...
Q: ಮಳೆ ಕಡಿಮೆ ಇದ್ದಾಗ ಭತ್ತ ಬೆಳೆಗೆ ಏನು ಮಾಡಬೇಕು
A: Spraying of 10% Bavistin @ 250ml/acre for 3 consecutive days in the morning and evening...
Limitations
Small training subset (600/69,670 available examples), 1 epoch — intended as a working proof of concept for the pipeline, not a production-ready model. Next iteration will scale to the full dataset with multi-epoch training.
License
Apache 2.0 (inherits from base model)