|
--- |
|
library_name: peft |
|
--- |
|
<a target="_blank" href="https://colab.research.google.com/github/szymonrucinski/finetune-llm/blob/main/pollama_Xb_inference.ipynb"> |
|
<img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/> |
|
</a> |
|
|
|
## Introduction |
|
Krakowiak-7B is a finetuned version of Meta's [Llama2](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf). It was trained on the modified and updated dataset originally created by [Chris Ociepa](https://huggingface.co/datasets/szymonindy/ociepa-raw-self-generated-instructions-pl) |
|
containing ~ 50K instructions. Making it one of the best and biggest available LLM's. |
|
Name [krakowiak](https://www.youtube.com/watch?v=OeQ6jYzt6cM) refers to one of the most popular and characteristic Polish folk dances, with its very lively, even wild, tempo, and long, easy strides, demonstrating spirited abandon and elegance at the same time. |
|
|
|
## How to test it? |
|
The model can be ran using the Huggingface library or in the browser using this [Google Colab](https://colab.research.google.com/drive/1IM7j57g9ZHj-Pw2EXGyacNuKHjvK3pIc?usp=sharing) |
|
|
|
## Training procedure |
|
Model was trained for 3 epochs, feel free [to read the report](https://api.wandb.ai/links/szymonindy/tkr343ad) |
|
|
|
The following `bitsandbytes` quantization config was used during training: |
|
- load_in_8bit: False |
|
- load_in_4bit: True |
|
- llm_int8_threshold: 6.0 |
|
- llm_int8_skip_modules: None |
|
- llm_int8_enable_fp32_cpu_offload: False |
|
- llm_int8_has_fp16_weight: False |
|
- bnb_4bit_quant_type: nf4 |
|
- bnb_4bit_use_double_quant: False |
|
- bnb_4bit_compute_dtype: bfloat16 |
|
|
|
The following `bitsandbytes` quantization config was used during training: |
|
- load_in_8bit: False |
|
- load_in_4bit: True |
|
- llm_int8_threshold: 6.0 |
|
- llm_int8_skip_modules: None |
|
- llm_int8_enable_fp32_cpu_offload: False |
|
- llm_int8_has_fp16_weight: False |
|
- bnb_4bit_quant_type: nf4 |
|
- bnb_4bit_use_double_quant: False |
|
- bnb_4bit_compute_dtype: bfloat16 |
|
### Framework versions |
|
|
|
- PEFT 0.4.0 |
|
|
|
- PEFT 0.4.0 |
|
|
