---
language:
- en
base_model: mikahniehaus/Tern
pipeline_tag: text2text-generation
tags:
- tinycode
- onnx
- python
- reasoning
---
# tinycode-narrator

Accepted TinyCode model, based on Tern 124M. Python code -> steps -> summary. The ONNX graph includes the snippet LoRA adapter; native inference needs both checkpoint.pt and snippet_adapter.pt. Functions use the unchanged function checkpoint.

## Files
- `model.onnx`: complete FP32 browser inference graph, no external tensor files.
- `manifest.json`: SHA-256, graph dimensions and inference settings.
- `checkpoint.pt`: native PyTorch model (custom Tern architecture, not Transformers AutoModel).
- `runtime/codenarrator/`: Python architecture, tokenizer, preprocessing and decoding implementation.
- `gpt2.tiktoken`: complete GPT-2 mergeable token ranks (base64 token + rank).
- `tinycode_config.json`: architecture and task settings.

Upload ALL contents of this folder to the ROOT of a public Hugging Face model repository.
Keep filenames unchanged. The website downloads `model.onnx` using
`https://huggingface.co/USERNAME/REPO/resolve/main/model.onnx`.
The native checkpoints are provided for reproducibility; they are not downloaded by the website.
No hosted inference server is needed. Weights execute on the visitor's CPU via ONNX Runtime Web.

Original pretraining model: https://huggingface.co/mikahniehaus/Tern .