rAIfle's picture
Create README.md
a0fb734 verified
```
e88 88e d8
d888 888b 8888 8888 ,"Y88b 888 8e d88
C8888 8888D 8888 8888 "8" 888 888 88b d88888
Y888 888P Y888 888P ,ee 888 888 888 888
"88 88" "88 88" "88 888 888 888 888
b
8b,
e88'Y88 d8 888
d888 'Y ,"Y88b 888,8, d88 ,e e, 888
C8888 "8" 888 888 " d88888 d88 88b 888
Y888 ,d ,ee 888 888 888 888 , 888
"88,d88 "88 888 888 888 "YeeP" 888
PROUDLY PRESENTS
```
# Llama-3-70B-Instruct-Storywriter-exl2-rpcal
Quantized using 200 samples of 8192 tokens from an RP-oriented [PIPPA](https://huggingface.co/datasets/royallab/PIPPA-cleaned) dataset.
Branches:
- `main` -- `measurement.json`
- `2.25b6h` -- 2.25bpw, 6bit lm_head
- `3.5b6h` -- 3.5bpw, 6bit lm_head
- `3.75b6h` -- 3.75bpw, 6bit lm_head
- `4.5b6h` -- 4.5bpw, 6bit lm_head
- `4.65b6h` -- 4.65bpw, 6bit lm_head
- `6b6h` -- 6bpw, 6bit lm_head
- `8b8h` -- 8bpw, 8bit lm_head
Original model link: [tdrussell/Llama-3-70B-Instruct-Storywriter](https://huggingface.co/tdrussell/Llama-3-70B-Instruct-Storywriter)
Original model README below.
-----
# Llama 3 70B Instruct Storywriter
Llama 3 70B Instruct, further finetuned on a dataset consisting of books in the fiction genre.
This was just an experiment, but it turned out well enough that I'm sharing it. The finetuning has caused a significant shift in the model's writing style, and seems to have made it more creative. There may be a slight decrease in overall intelligence.
Because this was trained on Instruct, you can use the normal Instruct chat formatting. It may also work well in raw completion mode.
## Training details
Trained on 4 4090s using [qlora-pipe](https://github.com/tdrussell/qlora-pipe).
Dataset consists of about 800 books in the fiction genre, totaling 570 MB of raw text.
Rank 64 QLoRA trained at 8192 sequence length.
### Evaluation metrics
<img src="https://i.imgur.com/sCMjix4.png" width="800" />
## Why no 8B?
I tried multiple times to train this on Llama 3 8B Instruct, using a variety of hyperparameters. It never worked well. The model took a huge hit to intelligence every time, to the point of being unusable. 70B fared much better. I don't know why, maybe 8B is just too small for this type of technique, and loses too much of the instruction-tuned smarts.