|
--- |
|
license: mit |
|
datasets: |
|
- kaitchup/ultrachat-100k-flattened |
|
--- |
|
|
|
--- |
|
library_name: peft |
|
--- |
|
## Training procedure |
|
|
|
Finetuned on ultrachat-100k-flattened dataset for 1 epoch, took around 40 hrs on A100 80BG |
|
|
|
|
|
The following `bitsandbytes` quantization config was used during training: |
|
- quant_method: bitsandbytes |
|
- 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: True |
|
- bnb_4bit_compute_dtype: float16 |
|
### Framework versions |
|
|
|
|
|
- PEFT 0.5.0 |
|
|
|
### Prompt |
|
Use the following for prompting |
|
|
|
prompt = "### Human: "+instruction+"### Assistant: " |
|
|
|
|
|
#### merged following the gist https://gist.github.com/ChrisHayduk/1a53463331f52dca205e55982baf9930 |
|
|
|
#### Guidance from https://kaitchup.substack.com/ |
|
|
|
|
|
#### Work supported by https://datacrunch.io/ |
