Tamil-800M-Instruct (experimental)

SFT chat version of tamil-800m-base — a from-scratch 800M Tamil model. Fine-tuned 2 epochs on Tamil instruction data (tamil-alpaca, tamil-alpaca-orca, Aya, ai4bharat indic-instruct).

Status: experimental. It responds fluently in Tamil but is factually unreliable at this scale — treat it as a research artifact and fluency demo, not an assistant. A larger version (1B/3B + RL) is in progress.

See the base model card for architecture, training details, and the BPB evaluation where the base model outperforms Sarvam-1 (2.5B) on Tamil compression.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("skyblanket/tamil-800m-instruct")
model = AutoModelForCausalLM.from_pretrained("skyblanket/tamil-800m-instruct", dtype="bfloat16", device_map="auto")
ids = tok("இந்தியாவின் தலைநகரம் என்ன?", return_tensors="pt").to(model.device)
print(tok.decode(model.generate(**ids, max_new_tokens=100, do_sample=True, temperature=0.7, top_p=0.9, repetition_penalty=1.2)[0]))

Personal research project by skyblanket.