|
--- |
|
library_name: peft |
|
--- |
|
|
|
Authors: Forrest Kim, Joe Damisch, He Shi |
|
|
|
|
|
## Training procedure |
|
|
|
A BERT classifier was trained on the Stereoset dataset (https://huggingface.co/datasets/stereoset). Only the racial and professional prompts were used for training the classifier due to covariate imbalance. |
|
|
|
LLaMA 2 7B chat (https://huggingface.co/meta-llama/Llama-2-7b-chat-hf) was used to prompt a random sample of the Stereoset dataset prompts, as well as, from a manually curated list of leading prompts (What is wrong with doctors?, Why are programmers so skinny?, etc.) |
|
|
|
The outcomes of these prompts were then classified using the BERT Classifier and the resulting 'unbiased' context, prompt pairings were used to fine-tune the LLaMA 2 chat model using the PEFT LORA library. |
|
|
|
PEFT: |
|
|
|
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: fp4 |
|
bnb_4bit_use_double_quant: False |
|
bnb_4bit_compute_dtype: float32 |
|
|
|
This ran on for 1000 steps -- 1 epochs on a stereotype dataset. Training took ~1 hour on a 4090. |
|
|
|
### Framework versions |
|
|
|
|
|
- PEFT 0.5.0.dev0 |
