Palmyra-Fin-70B-32K is a model built by Writer specifically to meet the needs of the financial industry. It is a leading LLM on financial benchmarks, outperforming other large language models in various financial tasks and evaluations.

Resources and Technical Documentation:

Specialized for Financial Applications

Palmyra-Fin-70B-32K is meticulously designed to meet the unique linguistic and knowledge demands of the finance and economics sectors. It has been fine-tuned on an extensive collection of high-quality financial data, ensuring it can comprehend and generate text with precise domain-specific accuracy and fluency.

Our system integrates a specialized internal finance dataset and a well-crafted fine-tuning recipe, making it highly adept at handling the specific needs of this field. Key components of our training pipeline include:

Specialized Dataset: Utilizing a proprietary internal finance dataset to enhance the model’s performance.
Fine-tuning approach: Custom financial instruction dataset (Writer in-house build)

Model Description

Developed by: Writer
Model type: Llama
Language(s) (NLP): English
Context window: 32,768 tokens
Parameters: 70 billion
Finetuned from model: Palmyra-X-004
License: Writer open model license

Intended Use

Intended Use Cases

Palmyra-Fin-70B-32K is intended for use in English for financial analysis, market trend prediction, risk assessment, financial report generation, and automated financial advice. It excels at answering questions from long financial documents, making it ideal for in-depth financial research and analysis.

Out-of-scope

Use in any manner that violates applicable laws or regulations (including trade compliance laws). Use in any other way that is prohibited by Writer’s Acceptable Use Policy and the Writer open model license. Use in languages other than English.

Note: Users should be aware that while the model is highly capable, it should not be used as the sole basis for making significant financial decisions.

Use with transformers

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
model_id = "Writer/Palmyra-Fin-70B-32K"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
    attn_implementation="flash_attention_2",
)
messages = [
    {
        "role": "system",
        "content": "You are a highly knowledgeable and experienced expert in the financial sector, possessing extensive knowledge and practical expertise in financial analysis, markets, investments, and economic principles.",
    },
    {
        "role": "user",
        "content": "Can you explain how central banks printing more money (quantitative easing) affects the stock market and how investors might react to it?",
    },
]
input_ids = tokenizer.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
)
gen_conf = {
    "max_new_tokens": 1024,
    "eos_token_id": tokenizer.eos_token_id,
    "temperature": 0.0,
    "top_p": 0.9,
}
with torch.inference_mode():
    output_id = model.generate(input_ids, **gen_conf)
output_text = tokenizer.decode(output_id[0][input_ids.shape[1] :])
print(output_text)

Evaluation Results

Palmyra-Fin-70B-32K outperforms other models on internal finance evaluations, achieving state-of-the-art results across various financial datasets. Its strong performance in tasks like financial document analysis, market trend prediction, and risk assessment underscores its effective grasp of financial knowledge.

Key Performance Metrics:

100% accuracy on needle-in-haystack tasks
Superior performance on internal finance evaluations compared to other models

Palmyra-Fin-70B-32K achieves 100% accuracy on needle-in-haystack tasks across its entire 32,768 token context window, demonstrating exceptional capability in precise information extraction from extensive financial documents.

CFA Level III Test

The new Palmyra-Fin-70B model passed the CFA Level III test with a 73%. It’s the first model to pass this exam, often called “the world’s hardest exam.” The average passing score is 60%, and less than half of all test takers pass. This is a big improvement from other models like GPT-4, which scored 33% on the exam.

Long-Fin-Eval Performance:

To further assess the model’s capabilities, we developed and conducted an evaluation using long-fin-eval, an internally created benchmark designed to simulate real-world financial use cases. This evaluation consists of samples containing long documents paired with high-quality question-answer sets. The model’s task is to generate responses based on the provided document and question, with the output evaluated by GPT-4 Turbo.

The long-fin-eval methodology assesses both the model’s information retrieval capabilities and its ability to engage in extended dialogue on complex financial topics. This approach provides insight into the model’s capacity to process and synthesize information from lengthy financial documents while maintaining coherent and contextually appropriate conversational output.

In this evaluation, Palmyra-Fin-70B-32K showed superior performance compared to both open-source and proprietary benchmark models. These results indicate the model’s effectiveness in addressing real-world financial applications that require both comprehensive understanding of extensive documents and the ability to articulate nuanced insights.

model name	long-fin-eval
Palmyra-Fin-70B-32K	9.04
Claude 3.5 Sonnet	9.02
Qwen-2 70B instruct	8.9
gpt-4o	8.72
palmyra-fin-56b	8.23
mixtral-8x7b	7.57

Financial Use Cases

Palmyra-Fin-70B-32K excels in analyzing and summarizing complex financial reports, market data, and economic indicators, extracting key information to generate concise, structured summaries. It helps enhance financial decision-making by performing advanced entity recognition, identifying key financial concepts such as market trends, economic indicators, and financial instruments from unstructured text.

By leveraging its deep understanding of financial terminology, the model enhances information retrieval, data analysis, and knowledge discovery from financial reports, research articles, and other economic sources. These capabilities support applications like investment analysis, risk management, and financial research.

Bias, Risks, and Limitations

Palmyra-Fin-70B-32K, despite leveraging high-quality data, may contain inaccuracies, biases, or misalignments and has not been rigorously evaluated in real-world financial settings.

It is advised not to use the model for direct financial decision-making or professional financial advice without human oversight. Instead, its use should be confined to research and analysis by qualified individuals who understand its limitations. Palmyra-Fin-70B-32K should not replace professional financial judgment, and adapting it for critical financial use would require extensive additional work, including thorough testing, regulatory compliance, bias mitigation, and human oversight. Always consult a qualified financial professional for personal financial needs.

Citation and Related Information

To cite this model:

@misc{Palmyra-Fin-70B-32k,
  author = {Writer Engineering team},
  title = {{Palmyra-Fin-70B-32k: a powerful LLM designed for Finance}},
  howpublished = {\url{https://dev.writer.com}},
  year = 2024,
  month = July 
}

Contact Hello@writer.com