Welcome to the exciting world of Sujet Finance 8B v0.1 – your go-to language model for all things finance! 💰 This state-of-the-art model is a fine-tuned version of the powerful LLAMA 3 model, meticulously trained on the comprehensive Sujet Finance Instruct-177k dataset. 📈

🎯 Fine-Tuning Focus

In this initial fine-tuning iteration, we've focused on three key financial tasks:

✅❌ Yes/No Questions
- Description: This task involves answering financial questions that require a simple "yes" or "no" response.
- Class Distribution:
  - Train Set: 5,265 "yes" examples, 5,302 "no" examples
  - Eval Set: 1,340 "yes" examples, 1,303 "no" examples
📂 Topic Classification
- Description: The model classifies financial texts into specific finance-related categories such as company news, markets, earnings, and more.
- Class Distribution:
  - Train Set: Balanced across 20 classes, with 29-40 examples per class
  - Eval Set: Varies across classes, ranging from 4 to 15 examples per class
😊😐😡 Sentiment Analysis
- Description: This task involves analyzing financial texts to categorize sentiments as positive, negative, neutral, bearish, or bullish.
- Class Distribution:
  - Train Set: 1,160 positive, 1,155 negative, 1,150 neutral, 1,133 bearish, and 1,185 bullish examples
  - Eval Set: 281 positive, 286 negative, 291 neutral, 308 bearish, and 256 bullish examples

Inference code

This model was finetuned using Unsloth. Please refer to their github repository and make sure you have it installed before using the model : Unsloth

python

from unsloth import FastLanguageModel


max_seq_length = 2048 
dtype = None 
load_in_4bit = False 


alpaca_prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
{}

### Input:
{}

### Response:
{}"""


model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "sujet-ai/Sujet-Finance-8B-v0.1",
    max_seq_length = max_seq_length,
    dtype = dtype,
    load_in_4bit = load_in_4bit,
    token = "your hf token here",
)


example = {
'system_prompt': 'You are a financial sentiment analysis expert. Your task is to analyze the sentiment expressed in the given financial text.Only reply with bearish, neutral, or bullish.',
'user_prompt': "Expedia's Problems Run Deeper Than SEO Headwinds",
'answer': 'bearish',
}


inputs = tokenizer(
                [alpaca_prompt.format(
                    example['system_prompt'],  # instruction
                    example['user_prompt'],  # input
                    "",  # output - leave this blank for generation!
                )],
                return_tensors="pt"
            ).to("cuda")
            
outputs = model.generate(**inputs, max_new_tokens=2048, use_cache=True, pad_token_id=tokenizer.eos_token_id)
output = tokenizer.batch_decode(outputs)[0]
response = output.split("### Response:")[1].strip()
print(response)

You can find more information about the dataset by clicking on this link : Sujet-Finance-Instruct-177k Dataset

Our model has been carefully trained to excel in these areas, providing accurate and insightful responses to your financial queries. 💡

🎓 Training Methodology

To ensure optimal performance, we've employed a balanced training approach. Our dataset preparation process strategically selects an equal number of examples from each subclass within the three focus tasks. This results in a well-rounded model that can handle a diverse range of financial questions and topics. 🧠

The final balanced training dataset consists of 17,036 examples, while the evaluation dataset contains 4,259 examples.

🔧 Model Specifications

Base Model: LLAMA 3 8B 🦙
Fine-Tuning Technique: LoRA (Low-Rank Adaptation)
- r = 16
- alpha = 32
Learning Rate: 2e-4 📈
Weight Decay: 0.01 🏋️‍♂️
Epochs: 1 🔄
Quantization: float16 for VLLM 🗜️

📊 Evaluation Results

We've put our model to the test, comparing its performance against the base LLAMA 3 model on our evaluation dataset. The results are impressive! 🏆

We consider a response correct if the true answer appears within the first 10 words generated by the model. This strict criterion ensures that our model not only provides accurate answers but also prioritizes the most relevant information. 🎯