picollm-v1

picoLLM Runtime Artifacts

This repo contains the inference-focused artifact set produced by a picollm/accelerated/speedrun.sh run.

Provenance

• picoLLM repo commit: 7c2c7d831050ca0ca64265b85a4ffc9ba97b2596
• Preferred chat source: sft
• Base checkpoint: {'root': 'base_checkpoints', 'model_tag': 'd24', 'step': 5568, 'latest_model': 'base_checkpoints/d24/model_005568.pt', 'latest_meta': 'base_checkpoints/d24/meta_005568.json'}
• SFT checkpoint: {'root': 'chatsft_checkpoints', 'model_tag': 'd24', 'step': 482, 'latest_model': 'chatsft_checkpoints/d24/model_000482.pt', 'latest_meta': 'chatsft_checkpoints/d24/meta_000482.json'}
• Run manifest: run_manifest.json
• Identity dataset: identity_conversations.jsonl

Intended Usage

• restore the bundle into PICOLLM_BASE_DIR
• run python -m picollm.accelerated.chat.cli -i sft
• run python -m picollm.accelerated.chat.web -i sft

Demo

• Hugging Face Space: montekkundan/picollm-v1-demo

Learn More

• Build notes and deployment walkthrough: How to Create a LLM from scratch and deploy it

Known Limitations

• picoLLM-native checkpoints only; this repo is not Transformers-native
• optimizer shards are intentionally excluded here
• use the paired archive dataset repo if you need fuller resume-training artifacts
• the model can hallucinate and should be treated as a smaller open model

Local Restore

git clone https://github.com/Montekkundan/llm
cd llm
uv sync --extra gpu
python scripts/restore_picollm_from_hf.py montekkundan/picollm-v1

If you already downloaded the repo contents into a custom artifact directory:

export PICOLLM_BASE_DIR=$PWD/artifacts/picollm
python -m picollm.accelerated.chat.cli -i sft