Authors: Ashvini Kumar Jindal, Pawan Kumar Rajpoot, Ankur Parikh, Akshita Sukhlecha

🤗 Hugging Face Announcement Blog: https://huggingface.co/blog/akjindal53244/llama31-storm8b

🚀Ollama: ollama run ajindal/llama3.1-storm:8b

Llama-3.1-Storm-8B-GGUF

This is the GGUF quantized version of Llama-3.1-Storm-8B, for use with llama.cpp. BF16 Model here

TL;DR

We present the Llama-3.1-Storm-8B model that outperforms Meta AI's Llama-3.1-8B-Instruct and Hermes-3-Llama-3.1-8B models significantly across diverse benchmarks as shown in the performance comparison plot in the next section. Our approach consists of three key steps:

1. Self-Curation: We applied two self-curation methods to select approximately 1 million high-quality examples from a pool of ~2.8 million open-source examples. Our curation criteria focused on educational value and difficulty level, using the same SLM for annotation instead of larger models (e.g. 70B, 405B).
2. Targeted fine-tuning: We performed Spectrum-based targeted fine-tuning over the Llama-3.1-8B-Instruct model. The Spectrum method accelerates training by selectively targeting layer modules based on their signal-to-noise ratio (SNR), and freezing the remaining modules. In our work, 50% of layers are frozen.
3. Model Merging: We merged our fine-tuned model with the Llama-Spark model using SLERP method. The merging method produces a blended model with characteristics smoothly interpolated from both parent models, ensuring the resultant model captures the essence of both its parents. Llama-3.1-Storm-8B improves Llama-3.1-8B-Instruct across 10 diverse benchmarks. These benchmarks cover areas such as instruction-following, knowledge-driven QA, reasoning, truthful answer generation, and function calling.

🏆 Introducing Llama-3.1-Storm-8B

Llama-3.1-Storm-8B builds upon the foundation of Llama-3.1-8B-Instruct, aiming to enhance both conversational and function calling capabilities within the 8B parameter model class.

As shown in the left subplot of the above figure, Llama-3.1-Storm-8B model improves Meta-Llama-3.1-8B-Instruct across various benchmarks - Instruction-following (IFEval), Knowledge-driven QA benchmarks (GPQA, MMLU-Pro), Reasoning (ARC-C, MuSR, BBH), Reduced Hallucinations (TruthfulQA), and Function-Calling (BFCL). This improvement is particularly significant for AI developers and enthusiasts who work with limited computational resources.

We also benchmarked our model with the recently published model Hermes-3-Llama-3.1-8B built on top of the Llama-3.1-8B-Instruct model. As shown in the right subplot of the above figure, Llama-3.1-Storm-8B outperforms Hermes-3-Llama-3.1-8B on 7 out of 9 benchmarks, with Hermes-3-Llama-3.1-8B surpassing Llama-3.1-Storm-8B on the MuSR benchmark and both models showing comparable performance on the BBH benchmark.

Llama-3.1-Storm-8B Model Strengths

Llama-3.1-Storm-8B is a powerful generalist model useful for diverse applications. We invite the AI community to explore Llama-3.1-Storm-8B and look forward to seeing how it will be utilized in various projects and applications.

Model Strength	Relevant Benchmarks
🎯 Improved Instruction Following	IFEval Strict (+3.93%)
🌐 Enhanced Knowledge Driven Question Answering	GPQA (+7.21%), MMLU-Pro (+0.55%), AGIEval (+3.77%)
🧠 Better Reasoning	ARC-C (+3.92%), MuSR (+2.77%), BBH (+1.67%), AGIEval (+3.77%)
🤖 Superior Agentic Capabilities	BFCL: Overall Acc (+7.92%), BFCL: AST Summary (+12.32%)
🚫 Reduced Hallucinations	TruthfulQA (+9%)

Note: All improvements are absolute gains over Meta-Llama-3.1-8B-Instruct.

Llama-3.1-Storm-8B Models

1. BF16: Llama-3.1-Storm-8B
2. ⚡ FP8: Llama-3.1-Storm-8B-FP8-Dynamic
3. ⚡ GGUF: Llama-3.1-Storm-8B-GGUF
4. 🚀 Ollama: ollama run ajindal/llama3.1-storm:8b

💻 How to Use GGUF Model

pip install llama-cpp-python

from huggingface_hub import hf_hub_download
from llama_cpp import Llama

## Download the GGUF model
model_name = "akjindal53244/Llama-3.1-Storm-8B-GGUF"
model_file = "Llama-3.1-Storm-8B.Q8_0.gguf" # this is the specific model file we'll use in this example. It's a 4-bit quant, but other levels of quantization are available in the model repo if preferred
model_path = hf_hub_download(model_name, filename=model_file)

## Instantiate model from downloaded file
llm = Llama(
    model_path=model_path,
    n_ctx=16000,    # Context length to use
    n_threads=32,   # Number of CPU threads to use
    n_gpu_layers=0  # Number of model layers to offload to GPU
)

generation_kwargs = {
    "max_tokens":200,
    "stop":["<|eot_id|>"],
    "echo":False, # Echo the prompt in the output
    "top_k":1 # Set this value > 1 for sampling decoding
}

prompt = "What is 2+2?"
res = llm(prompt, **generation_kwargs)
print(res["choices"][0]["text"])

Function Calling Example with Ollama

import ollama
tools = [{
      'type': 'function',
      'function': {
        'name': 'get_current_weather',
        'description': 'Get the current weather for a city',
        'parameters': {
          'type': 'object',
          'properties': {
            'city': {
              'type': 'string',
              'description': 'The name of the city',
            },
          },
          'required': ['city'],
        },
      },
    },
    {
      'type': 'function',
      'function': {
        'name': 'get_places_to_vist',
        'description': 'Get places to visit in a city',
        'parameters': {
          'type': 'object',
          'properties': {
            'city': {
              'type': 'string',
              'description': 'The name of the city',
            },
          },
          'required': ['city'],
        },
      },
    },
  ]
response = ollama.chat(
    model='ajindal/llama3.1-storm:8b',
    messages=[
        {'role': 'system', 'content': 'Do not answer to nay vulgar questions.'},
        {'role': 'user', 'content': 'What is the weather in Toronto and San Francisco?'}
        ],
    tools=tools
)
print(response['message'])  # Expected Response: {'role': 'assistant', 'content': "<tool_call>{'tool_name': 'get_current_weather', 'tool_arguments': {'city': 'Toronto'}}</tool_call>"}

Alignment Note

While Llama-3.1-Storm-8B did not undergo an explicit model alignment process, it may still retain some alignment properties inherited from the Meta-Llama-3.1-8B-Instruct model.

Cite Our Work

@misc {ashvini_kumar_jindal_2024,
    author       = { {Ashvini Kumar Jindal, Pawan Kumar Rajpoot, Ankur Parikh, Akshita Sukhlecha} },
    title        = { Llama-3.1-Storm-8B },
    year         = 2024,
    url          = { https://huggingface.co/akjindal53244/Llama-3.1-Storm-8B },
    doi          = { 10.57967/hf/2902 },
    publisher    = { Hugging Face }
}

Support Our Work

With 3 team-members spanned across 3 different time-zones, we have won NeurIPS LLM Efficiency Challenge 2023 and 4 other competitions in Finance and Arabic LLM space. We have also published SOTA mathematical reasoning model.

Llama-3.1-Storm-8B is our most valuable contribution so far towards the open-source community. We are committed in developing efficient generalist LLMs. We're seeking both computational resources and innovative collaborators to drive this initiative forward.