scratchq-110 / README.md
rsumit1234's picture
Describe training by time, not cost
96821db verified
|
Raw History Blame Contribute Delete
646 Bytes
---
title: nanochat-d20 (trained from scratch)
emoji: 💬
colorFrom: indigo
colorTo: green
sdk: gradio
sdk_version: 5.49.1
python_version: "3.12"
app_file: app.py
pinned: false
license: mit
models:
- sumitkClasses/nanochat-d20-chat
---
Chat with a 20-layer model trained from scratch (tokenizer, pretraining, SFT) with
[karpathy/nanochat](https://github.com/karpathy/nanochat) on a single H100 (~5 h pretraining + 47 min chat fine-tuning).
Weights and training details: [sumitkClasses/nanochat-d20-chat](https://huggingface.co/sumitkClasses/nanochat-d20-chat).
`nanochat/` is vendored from nanochat commit 92d63d4 (MIT, see LICENSE-nanochat).