|
--- |
|
library_name: peft |
|
base_model: openlm-research/open_llama_3b |
|
--- |
|
|
|
# OpenLLaMa 3B PersonaChat |
|
This is a LoRA finetune of [OpenLLaMa 3B](https://huggingface.co/openlm-research/open_llama_3b) on the [personachat-truecased](https://huggingface.co/datasets/bavard/personachat_truecased) dataset with 3 epochs of 500 steps. |
|
|
|
## Use |
|
Before using this model, you must first add these extra tokens: |
|
```python |
|
tokenizer.add_special_tokens({"additional_special_tokens": ["<|human|>", "<|bot|>", "<|endoftext|>"]}) |
|
model.resize_token_embeddings(len(tokenizer)) |
|
``` |
|
The model is finetuned with the format is as follows: |
|
``` |
|
Personality: |
|
- [...] |
|
- [...] |
|
<|human|>Hi there!<|endoftext|><|bot|>Hello!<|endoftext|> |
|
``` |
|
To use this model, you must first define the personalities. |
|
```python |
|
personalities = """Personality: |
|
- [...] |
|
- [...] |
|
""" |
|
``` |
|
Then, follow the format: |
|
```python |
|
user = input(">>> ") |
|
prompt = f"{personalities}<|human|>{user}<|endoftext|><|bot|>" |
|
``` |
|
|
|
## Naming Format |
|
[model name]-finetuned-[dataset]-e[number of epochs]-s[number of steps] |
|
|
|
## Training procedure |
|
|
|
|
|
The following `bitsandbytes` quantization config was used during training: |
|
- load_in_8bit: True |
|
- load_in_4bit: False |
|
- 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 |
|
### Framework versions |
|
|
|
|
|
- PEFT 0.4.0.dev0 |
|
|
