tara10m-sft-v1-2k

Model Details

• Architecture: Llama-style decoder-only causal LM
• Parameters: 10,390,784
• Layers: 6
• Hidden size: 256
• Attention heads: 4
• Vocabulary: 16,000 SentencePiece tokens
• Context length: 1,024
• Base checkpoint: Tara10M Colab base checkpoint
• SFT data: 2,000 cleaned synthetic Burmese-English instruction examples

Intended Use

Best test areas:

• short English to Burmese translation
• short Burmese to English translation
• simple Burmese rewrite
• travel phrasebook style prompts

Prompt format:

text
Copy code
Instruction: Translate this sentence to Burmese.
Input: I will go home tomorrow.
Response:

Limitations

This model is very small and still repeats, drifts, and gives wrong translations. It should be used for experimentation only.

Known issues:

• weak factual reliability
• repeated phrases
• mixed Burmese/English output
• poor long-form generation
• not suitable for safety-critical use

Training Summary

text
Copy code
SFT examples: 2,000
Train examples: 1,800
Validation examples: 200
Max steps: 400
Learning rate: 2e-5
Best validation loss: 2.4228

Files

• model.safetensors
• config.json
• tokenizer.model
• tokenizer_config.json
• special_tokens_map.json