README.md

---
license: bigscience-bloom-rail-1.0
datasets:
- c4
language:
- en
library_name: transformers
tags:
- causal-lm
- gpt-j
---
6.7m (6,700,128) param GPT-J model.
```
n_positions - 128
n_embd - 64
n_layer - 4
n_head - 8
rotary_dim - 64
tokenizer - gpt-j
```

First, trained on 4,194,304 samples from the [c4](https://hf.co/datasets/c4) dataset, at a length of 128 tokens each, that comes out to 536,870,912 (0.53B) tokens seen during training. A batch size of 16 with 128 gradient accumulation steps was used, making the effective batch size 2048. A cosine learning rate schedule was used starting at 1e-3.

Second, the same settings, with double the accumulation steps, were used on [EleutherAI/the_pile_deduplicated](https://hf.co/EleutherAI/the_pile_deduplicated) with a learning rate of 1e-5 with a linearly decreasing learning rate schedule after the training on c4 was done. (for a total of ~1.06B tokens seen)


it's like this because i forgot to train on the pile in the first place and i wasn't going to let somewhere around 13 hours of gpu time go to waste I couldn't play minecraft that entire time