Model Details
This Repository is a 4-bit quantized version of Dorna-Llama3-8B-Instruct model for efficient memory usage. Dorna model is a decoder-only model, specifically trained/fine-tuned on Persian data. Flash Attention 2 is also integrated for faster inference.
Benefits
- Reduced Memory Usage: 4-bit quantization lowers memory requirements.
- Faster Inference: Flash Attention 2 speeds up processing.
- Easy Deployment: No need for additional libraries like LlamaCPP or Candle.
- Ready to Use: Compatible with Langchain, Haystack, LlamaIndex 2, and more.
- Google Colab Friendly: Can run on Google Colab free tier with T4 GPU (less than 15 GB of GPU RAM).
How to use
You can run conversational inference using the Transformers Auto classes with the generate() function. Let's look at an example.
import torch
import transformers
from transformers import AutoTokenizer, AutoModelForCausalLM
model_path = "amirMohammadi/Dorna-Llama3-8B-Instruct-Quantized4Bit"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(
model_path,
torch_dtype=torch.bfloat16,
device_map="auto",
)
messages = [
{"role": "system",
"content": "You are a helpful Persian assistant. Please answer questions in the asked language."},
{"role": "user", "content": "اصفهان بزرگ تر است یا قم؟"},
]
input_ids = tokenizer.apply_chat_template(
messages,
add_generation_prompt=True,
return_tensors="pt"
).to(model.device)
terminators = [
tokenizer.eos_token_id,
tokenizer.convert_tokens_to_ids("<|eot_id|>")
]
outputs = model.generate(
input_ids,
max_new_tokens=256,
eos_token_id=terminators,
do_sample=True,
temperature=0.6,
top_p=0.9,
)
response = outputs[0][input_ids.shape[-1]:]
print(tokenizer.decode(response, skip_special_tokens=True))
Evaluation of Non-Quantized version
This model is evaluated on questions across various tasks, including Boolean Questions, Code Generation, Long Response, Math, News QA, Paraphrasing, General Knowledge, and Summarization. Most categories typically have two main difficulty levels: Hard and Easy.
Both human evaluation and automatic evaluation (with GPT-4 as the judge) are performed.
In both tables, Dorna-8B-it is used as an abbreviated form of Dorna-Llama3-8B-Instruct.
Overall human evaluation results are as follows:
| Model Pairs | Parameters | Win % | Lose % | Tie % |
|---|---|---|---|---|
| Dorna-8B-it vs. Meta-Llama-3-8B-Instruct | 8B | 36.94 | 17.39 | 45.67 |
| Dorna-8B-it vs. GPT 3.5 turbo-1106 | N.A. | 32.01 | 26.94 | 41.05 |
| Dorna-8B-it vs. Persian Mind | 7B | 55.77 | 10.49 | 33.74 |
Category-based human evaluation results are as follows:
Win/Lose/Tie % is reported for each category.
Automatic evaluation results are as follows:
| Model Pairs | Parameters | Overall Win Rate % | Easy Win Rate % | Hard Win Rate % |
|---|---|---|---|---|
| Dorna-8B-it vs. Llama 3 base | 8B | 58.96 | 56.00 | 64.49 |
| Dorna-8B-it vs. Part Mistral | 7B | 77.20 | 73.00 | 85.05 |
| Dorna-8B-it vs. Persian Mind | 7B | 90.88 | 87.50 | 97.20 |
| Dorna-8B-it vs. Neuraorca Gemma 7b | 7B | 86.32 | 86.50 | 85.98 |
| Dorna-8B-it vs. Maral 7b | 7B | 97.39 | 97.00 | 98.13 |
| Dorna-8B-it vs. PersianLlama 7b | 7B | 98.70 | 98.00 | 100.00 |
| Dorna-8B-it vs. Aya-23-8B | 8B | 52.77 | 56.50 | 45.79 |
| Dorna-8B-it vs. Aya-23-35B | 35B | 45.93 | 54.00 | 30.84 |
| Dorna-8B-it vs. Command R | 35B | 58.63 | 61.00 | 54.21 |
Contact us
If you have any questions regarding this model, you can reach us via the community on Hugging Face.