metadata
library_name: transformers
license: apache-2.0
base_model: Qwen/Qwen3-4B-Base
tags:
- llm
- indic
model-index:
- name: Hex-1
results: []
language:
- hi
- te
- ta
- ml
- kn
Hex-1
This model is a fine-tuned version of Qwen/Qwen3-4B-Base on the alpaca_hindi, the alpaca_tamil, the alpaca_telugu and the alpaca_kannada datasets.
Model description
Hex-1 is a 4-billion parameter language model specifically optimized for Indian languages. It supports five major Indian languages, including Hindi, Kannada, Telugu, Tamil and Malayalam. When benchmarked against leading models like Gemma-2B, LLaMA-3.2-3B, and Sarvam-1, Hex1 delivers best-in-class performance in all five supported languages on MMLU dataset.
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- distributed_type: multi-GPU
- num_devices: 7
- total_train_batch_size: 56
- total_eval_batch_size: 56
- optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: cosine
- lr_scheduler_warmup_ratio: 0.1
- num_epochs: 3.0
Training results
Performance Comparison on ARCC Dataset
| Benchmark | Gemma-2-2B | Llama-3.2-3B | Llama-3.1-8B | Sarvam-1 | Hex-1 |
|---|---|---|---|---|---|
| arcc_hi | 37.57 | 49.13 | 56.17 | 60.00 | 36.68 |
| arcc_ta | 32.78 | 34.70 | 44.78 | 57.04 | 38.65 |
| arcc_te | 30.00 | 34.09 | 43.04 | 59.39 | 37.96 |
| arcc_kn | 29.22 | 36.43 | 44.70 | 57.04 | 38.31 |
| arcc_ml | 29.91 | 33.22 | 46.78 | 58.96 | 29.60 |
Framework versions
- Transformers 4.51.3
- Pytorch 2.7.0+cu126
- Datasets 3.5.0
- Tokenizers 0.21.1