sarvam1-hinglish-tn-lora

Why

The base sarvam-1 is a base (non-instruct) model and cannot be reliably prompted into TN (12-shot ICL scores 49.9% WER — worse than rules). This adapter is fine-tuned on a synthetic, correct-by-construction code-mixed corpus.

Results (held-out 40-sentence labeled set)

Usage

python
Copy code
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("sarvamai/sarvam-1")
m = AutoModelForCausalLM.from_pretrained("sarvamai/sarvam-1")
m = PeftModel.from_pretrained(m, "AK04-IXR/sarvam1-hinglish-tn-lora")

prompt = "Input: Mera flight ticket PNR-8392 hai, aur departure 4:30 PM ko hai.\nOutput:"
ids = tok(prompt, return_tensors="pt").to(m.device)
out = m.generate(**ids, max_new_tokens=96, do_sample=False)
print(tok.decode(out[0][ids['input_ids'].shape[1]:], skip_special_tokens=True))
# -> Mera flight ticket pee-en-aar eight three nine two hai, aur departure four thirty pee-em ko hai.

Training

LoRA (r=16, α=32, all attn+MLP projections; 0.94% of params) on ~8k synthetic pairs, 3 epochs, bf16, on a single A100. See the GitHub repo for the data generator, trainer, and evaluation harness.

Limitations

Trained on synthetic data, so it follows the project's normalization conventions; the held-out test set is small (40 sentences) — treat the headline number as indicative and see the per-category breakdown in the repo.