HeliosLM — HF distribution package (v5.16-era snapshot)

HeliosLM is a from-scratch, pure-PyTorch reference implementation of a DeepSeek-V3-style LLM stack, built to be read, verified on CPU, and hacked on. Correctness-first research line (agentic LM with certified-confidence routing).

• Full source: https://github.com/tonythetiger168/helioslm (main @ v5.31)
• Current checkpoints: https://huggingface.co/chienhsinlin/helioslm

What this repo contains

helioslm_hf_package_v5.16.zip (31.5 MB) — a self-contained HF-ready snapshot of the toy reference model as of v5.16: toy_v5.13.pt (8.5M char-level), config_lite.json, generation_example.py, a model card, a demo Space, and upload tooling.

Verified working (2026-09-25)

toy_v5.13.pt loads and generates with the current GitHub main (v5.31) code — greedy output on prompt "HeliosLM": 'HeliosLMv5Config.\n...'. It is a TOY checkpoint (trained on the repo's own source, val_loss 2.41): output is illustrative, not useful capability.

How to use (NOT a transformers model)

No config.json/safetensors; AutoModel.from_pretrained will not work — that is intentional.

git clone https://github.com/tonythetiger168/helioslm
cd helioslm && pip install torch
wget https://huggingface.co/chienhsinlin/helios/resolve/main/helioslm_hf_package_v5.16.zip
unzip helioslm_hf_package_v5.16.zip
cd hf_helioslm && python generation_example.py

Status

v5.16-era snapshot; the line has since shipped the agent layer (v5.23–29), chat capability (v5.30.2: text-channel mode-choice 11/20), and five frontier-gap reference modules (v5.31: mHC, compressed attention, long-horizon env, think+experience reuse, async GRPO). A refreshed package cut from main is planned; the snapshot's checkpoint remains compatible with the current codebase (verified above).