22 lines
795 B
Markdown
22 lines
795 B
Markdown
|
---
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
tags:
|
||
|
|
- text generation
|
||
|
|
- pytorch
|
||
|
|
- causal-lm
|
||
|
|
- gpt_neox
|
||
|
|
license: mit
|
||
|
|
datasets:
|
||
|
|
- hoskinson-center/proof-pile
|
||
|
|
---
|
||
|
|
|
||
|
|
# ProofGPT-v0.1
|
||
|
|
|
||
|
|
# Model Description
|
||
|
|
ProofGPT-v0.1 is a 1.3B parameter language model based on the GPT-NeoX architecture and trained on the [proof-pile](https://huggingface.co/datasets/hoskinson-center/proof-pile) (v1.1).
|
||
|
|
We initiailized training with pythia-1.3b weights, a precursor to the [pythia-1.4b](https://huggingface.co/EleutherAI/pythia-1.3b) model that has roughly equivalent performance.
|
||
|
|
|
||
|
|
Detailed evaluations coming soon :)
|
||
|
|
|
||
|
|
**Note**: Commit `3bcdc4e` replaced the weights with a model trained on proof-pile v1.1, as opposed to previous commits which were trained on v1.0. Commit `9695b51` updated the tokenizer to have bos, eos, and unk tokens.
|