EOS-llm identity data
eos_llm_identity.jsonl contains short conversation turns for teaching the model to identify as EOS-llm:
- Name: EOS-llm
- Developed by: AI team @ Safire
- Head of AI at Safire: Swaroop Kallakuri
- What it does: Language model built for energy efficiency, designed to run on a laptop
Format: one JSON array per line, each with alternating user / assistant messages (same as identity_conversations.jsonl).
Fine-tuning with this data
SFT automatically includes this file when present. From repo root:
Single GPU:
python -m scripts.chat_sft --run eos-llm-sft
8 GPUs (e.g. A100):
torchrun --standalone --nproc_per_node=8 -m scripts.chat_sft -- --device-batch-size=8 --run eos-llm-sft
After SFT, test in chat CLI:
python -m scripts.chat_cli -i sft
# Ask: "What's your name?" / "Who developed you?" / "What do you do?"
Or run chat eval and then start the web UI:
python -m scripts.chat_eval -i sft -a ARC-Easy
python -m scripts.chat_web -i sft
To add more EOS-llm Q&A pairs, append lines to eos_llm_identity.jsonl in the same JSONL format (one conversation per line).