bharadwajswarna/gemma-2b-sft-telugu

Baby Nandi

Baby Nandi (part of the Nandi series of Telugu LLMs) is a Telugu Instruction Tuned Version of Gemma 2B, part of an attempt to develop smaller and efficient Indic LLMs, useful for practical purposes. It beats the original gemma-2b overall, but still is behind the latest gemma-2b-1.1-it.

🏆 Benchmarks

Model	Average	AGIEval	GPT4All	TruthfulQA	Bigbench
bharadwajswarna/gemma-2b-sft-telugu📄	38.99	21.53	55.56	48.33	30.56
google/gemma-2b-it 📄	36.1	23.76	43.6	47.64	29.41
google/gemma-2b 📄	34.26	22.7	43.35	39.96	31.03

Training Process & Datasets :

1. Gemma 2b Base model has been further pretrained on a part of AI4Bharat Sangraha dataset (280k Telugu Samples).
2. SFT on a mix of Telugu Alpaca + Telugu GPTeacher from Telugu LLM Labs and English Alpaca

You can find the link to this model here : Gemma-2b-Telugu-Base-Model

Training Duration :

1. Pretraining for 6 epochs, nearly 35 hours (This might not be enough)
2. SFT for 3 epochs

Inference Prompt Template:

"""
### Instruction:
{}

### Input:
{}

### Response:
{}
"""

Developer : Bharadwaj Swarna
You can reach out to me for any questions/suggestions/collaborations.