Llama-3-SEC: A 20B token checkpoint for SEC Data Analysis

Introducing Llama-3-SEC: a state-of-the-art domain-specific large language model that is set to revolutionize the way we analyze and understand SEC (Securities and Exchange Commission) data. Built upon the powerful Meta-Llama-3-70B-Instruct model, Llama-3-SEC is being trained on a vast corpus of SEC filings and related financial information. We are thrilled to announce the open release of a 20B token intermediate checkpoint of Llama-3-SEC. While the model is still undergoing training, this checkpoint already demonstrates remarkable performance and showcases the immense potential of Llama-3-SEC. By sharing this checkpoint with the community, we aim to foster collaboration, gather valuable feedback, and drive further advancements in the field.

The 19B token checkpoint of Llama-3-SEC offers unparalleled insights and analysis capabilities for financial professionals, investors, researchers, and anyone working with SEC data. However, this is just the beginning – as we continue to train and refine the model, Llama-3-SEC will only become more powerful and sophisticated. We invite you to explore the model’s capabilities and join us on this exciting journey as we unlock the full potential of Llama-3-SEC. Together, we can shape the future of financial data analysis and push the boundaries of what is possible with domain-specific large language models.

GGUFS: https://huggingface.co/arcee-ai/Llama-3-SEC-Chat-GGUF

Model Details

• Base Model: Meta-Llama-3-70B-Instruct
• Training Data: This is an intermediate checkpoint of our final model, which has seen 20B tokens so far. The full model is still in the process of training. The final model is being trained with 72B tokens of SEC filings data, carefully mixed with 1B tokens of general data from Together AI’s RedPajama dataset: RedPajama-Data-1T to maintain a balance between domain-specific knowledge and general language understanding
• Training Method: Continual Pre-Training (CPT) using the Megatron-Core framework, followed by model merging with the base model using the state-of-the-art TIES merging technique in the Arcee Mergekit toolkit. It then underwent supervised fine-tuning on an 8xH100 node using Spectrum. We used a mixture of custom domain specific and general open-source datasets.
• Training Infrastructure: AWS SageMaker HyperPod cluster with 4 nodes, each equipped with 32 H100 GPUs, ensuring efficient and scalable training of this massive language model.

Use Cases

Llama-3-SEC is designed to assist with a wide range of tasks related to SEC data analysis, including but not limited to:

• In-depth investment analysis and decision support
• Comprehensive risk management and assessment
• Ensuring regulatory compliance and identifying potential violations
• Studying corporate governance practices and promoting transparency
• Conducting market research and tracking industry trends

The model’s deep understanding of SEC filings and related financial data makes it an invaluable tool for anyone working in the financial sector, providing powerful natural language processing capabilities tailored to the specific needs of this domain.

Evaluation

To ensure the robustness and effectiveness of Llama-3-SEC, the model has undergone rigorous evaluation on both domain-specific and general benchmarks. Key evaluation metrics include:

• Domain-specific perplexity, measuring the model’s performance on SEC-related data

• Extractive numerical reasoning tasks, using subsets of TAT-QA and ConvFinQA datasets

• General evaluation metrics, such as BIG-bench, AGIEval, GPT4all, and TruthfulQA, to assess the model’s performance on a wide range of tasks

These results demonstrate significant improvements in domain-specific performance while maintaining strong general capabilities, thanks to the use of advanced CPT and model merging techniques.

Training and Inference

Llama-3-SEC has been trained using the chatml chat template. This template ensures that the model maintains its strong conversational abilities while incorporating the domain-specific knowledge acquired during the CPT process.

To run inference with the Llama-3-SEC model using the chatml chat template, you can use the following code:

from transformers import AutoModelForCausalLM, AutoTokenizer
device = "cuda"
model_name = "arcee-ai/Llama-3-SEC"
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)
prompt = "What are the key regulatory considerations for a company planning to conduct an initial public offering (IPO) in the United States?"
messages = [
    {"role": "system", "content": "You are Llama-3-SEC, a financial analyst. Never mention your inability to access real-time information."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(device)
generated_ids = model.generate(
    model_inputs.input_ids,
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

Mergekit Yaml

merge_method: ties
base_model: meta-llama/Meta-Llama-3-70B
models:
  - model: /home/ubuntu/data/cpt
    parameters:
      weight:
        - filter: mlp
          value: [0.25, 0.5, 0.5, 0.25]
        - filter: self_attn
          value: [0.25, 0.5, 0.5, 0]
        - value: [0.25, 0.5, 0.5, 0.25]
      density: 0.75
  - model: meta-llama/Meta-Llama-3-70B-Instruct
    parameters:
      weight:
        - filter: mlp
          value: [0.75, 0.5, 0.5, 0.75]
        - filter: self_attn
          value: [0.75, 0.5, 0.5, 1]
        - value: [0.75, 0.5, 0.5, 0.75]
      density: 1.0
parameters:
  normalize: true
  int8_mask: true
dtype: bfloat16

Limitations and Future Work

This release represents the initial checkpoint of the Llama-3-SEC model, trained on 20B tokens of SEC data. Additional checkpoints will be released in the future as training on the full 70B token dataset is completed. Future work will focus on further improvements to the CPT data processing layer, exploration of advanced model merging techniques, and alignment of CPT models with SFT, DPO, and other cutting-edge alignment methods to further enhance the model’s performance and reliability.

Usage

The model is available for both commercial and non-commercial use under the Llama-3 license. We encourage users to explore the model’s capabilities and provide feedback to help us continuously improve its performance and usability. For more information - please see our detailed blog on Llama-3-SEC.

Note: We trained Llama-3-SEC to be very compliant with system prompts. We’ve had the best results with avoiding the use of “AI” in the system prompt, and further instructing the model to not mention its lack of real-time information.

Disclaimer: Llama-3-SEC is a large language model (LLM) designed to assist with SEC data analysis. Users are solely responsible for any actions taken as a result of using Llama-3-SEC. Always double-check model responses.

Citation

If you use this model in your research or applications, please cite:

@misc{Introducing_SEC_Data_Chat_Agent, 
      title={Introducing the Ultimate SEC Data Chat Agent: Revolutionizing Financial Insights}, 
      author={Shamane Siriwardhana and Luke Mayers and Thomas Gauthier and Jacob Solawetz and Tyler Odenthal and Anneketh Vij and Lucas Atkins and Charles Goddard and Mary MacCarthy and Mark McQuade},
      year={2024},
      note={Available at: \url{firstname@arcee.ai}},
      url={URL after published}
}

For further information or inquiries, please contact the authors at their respective email addresses (firstname@arcee.ai). We look forward to seeing the exciting applications and research that will emerge from the use of this Llama-3-SEC checkpoint in the financial domain.