yacdev's picture
Upload fine-tuned SomaliLM-150M-Instruct (100k Native SFT)
4ad9414 verified
|
Raw History Blame Contribute Delete
848 Bytes
metadata
language:
  - so
tags:
  - somali
  - causal-lm
  - sft
  - conversational
license: apache-2.0
datasets:
  - yacdev/somali-100k-native-conversations

πŸ‡ΈπŸ‡΄ SomaliLM-150M-Instruct (100K SFT)

SomaliLM-150M-Instruct is a lightweight, fluent Somali conversational language model fine-tuned on 100,000 pristine multi-turn native Somali conversations across 20 real-world domains.

πŸ“Š Specifications:

  • Parameters: ~150 Million
  • Training Dataset: 100,000 native Somali dialogues (579,556 dialogue turns, 17.5M words)
  • Format: ChatML (<|im_start|>user / assistant<|im_end|>)
  • Loss Strategy: Assistant Token Masking (100% loss computed on assistant responses)

πŸ’¬ Prompt Format (ChatML):

<|im_start|>user
Assalaamu calaykum walaal, sideen ku bilaabaa barashada Python-ka?
<|im_end|>
<|im_start|>assistant