|
--- |
|
license: other |
|
license_name: yi-license |
|
license_link: https://huggingface.co/01-ai/Yi-34B/blob/main/LICENSE |
|
language: |
|
- en |
|
library_name: transformers |
|
pipeline_tag: text-generation |
|
--- |
|
|
|
**NousResearch/Nous-Capybara-34B**, **migtissera/Tess-M-v1.2** and **migtissera/Tess-M-v1.3** merged with a new, experimental implementation of "dare ties" via mergekit. See: |
|
|
|
> Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch |
|
|
|
https://github.com/yule-BUAA/MergeLM |
|
|
|
https://github.com/cg123/mergekit/tree/dare-tokenizer |
|
|
|
It was quantized with exllamav2 on 200 rows (400K tokens) on a long Vicuna format chat, a single sci fi story and a single fantasy story. This should hopefully yield better chat performance than the default wikitext quantization. |
|
|
|
Quantized to 4bpw, enough for **~45K context on a 24GB GPU.** |
|
*** |
|
|
|
Merged with the following config, and the tokenizer from Yi Llamafied: |
|
``` |
|
models: |
|
- model: /home/alpha/Storage/Models/Raw/larryvrh_Yi-34B-200K-Llamafied |
|
# no parameters necessary for base model |
|
- model: /home/alpha/Storage/Models/Raw/migtissera_Tess-M-v1.3 |
|
parameters: |
|
weight: 0.50 |
|
density: 0.56 |
|
- model: /home/alpha/Storage/Models/Raw/migtissera_Tess-M-v1.2 |
|
parameters: |
|
weight: 0.20 |
|
density: 0.50 |
|
- model: /home/alpha/Storage/Models/Raw/Nous-Capybara-34B |
|
parameters: |
|
weight: 0.50 |
|
density: 0.56 |
|
merge_method: dare_ties |
|
base_model: /home/alpha/Storage/Models/Raw/larryvrh_Yi-34B-200K-Llamafied |
|
parameters: |
|
int8_mask: true |
|
dtype: bfloat16 |
|
``` |
|
|
|
Tess 1.2 (at a low weight) and 1.3 were used because, according to the trainer, they were trained on different datasets: https://migel.substack.com/p/learnings-from-training-tess |
|
|
|
I chose not to include other finetunes, such as Dolphin, because they aren't trained on the 200K base. If any other 200K finetunes pop up, let me know. |
|
|
|
*** |
|
|
|
|
|
First exllama quantization pass, on 80 rows so it will fit in memory: |
|
``` |
|
python convert.py --in_dir /home/alpha/FastModels/Capybara-Tess-34B-200K-DARE -o /home/alpha/FastModels/scratch -om /home/alpha/FastModels/capytess13mes.json --cal_dataset /home/alpha/Documents/smol.parquet -l 2048 -r 80 -ml 2048 -mr 40 -gr 40 -ss 4096 -nr -b 4.0 -hb 6 |
|
``` |
|
|
|
Second exllama quantization pass. 200 rows: |
|
``` |
|
python convert.py --in_dir /home/alpha/FastModels/Capybara-Tess-34B-200K-DARE -o /home/alpha/FastModels/scratch -m /home/alpha/FastModels/capytess13mes.json --cal_dataset /home/alpha/Documents/medium.parquet -l 2048 -r 200 -ml 2048 -mr 40 -gr 200 -ss 4096 -b 4.0 -hb 6 -cf /home/alpha/FastModels/Capybara-Tess-34B-200K-DARE-exl2-4bpw-fiction -nr |
|
``` |
|
|
|
|
|
*** |
|
## Prompt template: Orca-Vicuna |
|
|
|
``` |
|
SYSTEM: {system_message} |
|
USER: {prompt} |
|
ASSISTANT: |
|
|
|
``` |
|
Being a Yi model, try disabling the BOS token and/or running a lower temperature with MinP if output doesn't seem right. |
|
|
|
Sometimes the model "spells out" the stop token as `</s>` like Capybara, so you may need to add `</s>` as an additional stopping condition. |
|
|
|
*** |
|
|
|
Credits: |
|
|
|
https://github.com/cg123/mergekit/tree/dare-tokenizer |
|
|
|
https://huggingface.co/NousResearch/Nous-Capybara-34B/ |
|
|
|
https://huggingface.co/migtissera/Tess-M-v1.2 |
|
|
|
https://huggingface.co/migtissera/Tess-M-v1.3 |
|
|
|
https://huggingface.co/larryvrh/Yi-34B-200K-Llamafied |
|
|
|
https://huggingface.co/01-ai/Yi-34B-200K |
