EpistemeAI2/Fireball-Meta-Llama-3.1-8B-Instruct-Agent-0.003-128K-code-math

Agent LLama

Experimental and revolutionary fine-tune technique to allow LLama 3.1 8B to be agentic in math, coding and other applications. It has some build-in agent features:

• search
• calculator
• ReAct. Synergizing Reasoning and Acting in Language Models
  • fine tuned ReAct for better responses

Other noticable features:

• Self learning using unsloth. (in progress)
• can be used in RAG applications
• Memory. please use Langchain memory , section Message persistence

It is perfectly use for Langchain or LLamaIndex and Ollama.

• Check out Ollama Llama Agent in out github page,
• Context Window: 128K

Agent LLama series

• For coding agent, please go to EpistemeAI/Fireball-Meta-Llama-3.1-8B-Instruct-Agent-0.003-128K-code

Installation

pip install --upgrade "transformers>=4.43.2"

Developers can easily integrate EpistemeAI/Fireball-Meta-Llama-3.1-8B-Instruct-Agent-0.003-128K into their projects using popular libraries like Transformers and vLLM. The following sections illustrate the usage with simple hands-on examples:

Optional: to use build in tool, please add to system prompt: "Environment: ipython. Tools: brave_search, wolfram_alpha. Cutting Knowledge Date: December 2023. Today Date: 27 Auguest 2024\n"

ToT - Tree of Thought

• Use system prompt:

"Imagine three different experts are answering this question.
All experts will write down 1 step of their thinking,
then share it with the group.
Then all experts will go on to the next step, etc.
If any expert realises they're wrong at any point then they leave.
The question is..."

ReAct

example from langchain agent - langchain React agent

• Use system prompt:

"""
Answer the following questions as best you can. You have access to the following tools:

            {tools}

            Use the following format:

            Question: the input question you must answer
            Thought: you should always think about what to do
            Action: the action to take, should be one of [{tool_names}]
            Action Input: the input to the action
            Observation: the result of the action
            ... (this Thought/Action/Action Input/Observation can repeat N times)
            Thought: I now know the final answer
            Final Answer: the final answer to the original input question

            Begin!

            Question: {input}
            Thought:{agent_scratchpad}
"""

Conversational Use-case

Use with Transformers

Using transformers.pipeline() API , best use for 4bit for fast response.

import transformers
import torch
from langchain_community.llms import HuggingFaceEndpoint
from langchain_community.chat_models.huggingface import ChatHuggingFace

from transformers import BitsAndBytesConfig

quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype="float16",
    bnb_4bit_use_double_quant=True,
)

model_id = "EpistemeAI2/Fireball-Meta-Llama-3.1-8B-Instruct-Agent-0.003-128K-code-math"
pipeline = transformers.pipeline(
    "text-generation",
    model=model_id,
    model_kwargs={"quantization_config": quantization_config}, #for fast response. For full 16bit inference, remove this code.
    device_map="auto",
)
messages = [
    {"role": "system", "content":  """
    Environment: ipython. Tools: brave_search, wolfram_alpha. Cutting Knowledge Date: December 2023. Today Date: 4 October 2024\n
    You are a coding assistant with expert with everything\n
    Ensure any code you provide can be executed \n
    with all required imports and variables defined. List the imports.  Structure your answer with a description of the code solution. \n
    write only the code. do not print anything else.\n
    debug code if error occurs. \n
    Here is the user question: {question}
    """},
    {"role": "user", "content": "Create a bar plot showing the market capitalization of the top 7 publicly listed companies using matplotlib"}
]
outputs = pipeline(messages, max_new_tokens=128, do_sample=True, temperature=0.01, top_k=100, top_p=0.95)
print(outputs[0]["generated_text"][-1])  

Example:

Please go to Colab for sample of the code using Langchain Colab

Unsloth Fast

%%capture
# Installs Unsloth, Xformers (Flash Attention) and all other packages!
!pip install unsloth
# Get latest Unsloth
!pip install --upgrade --no-deps "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
!pip install langchain_experimental

from unsloth import FastLanguageModel
from google.colab import userdata


# 4bit pre quantized models we support for 4x faster downloading + no OOMs.
fourbit_models = [
    "unsloth/mistral-7b-instruct-v0.2-bnb-4bit",
    "unsloth/gemma-7b-it-bnb-4bit",
] # More models at https://huggingface.co./unsloth

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "EpistemeAI2/Fireball-Meta-Llama-3.1-8B-Instruct-Agent-0.003-128K-code-math",
    max_seq_length = 128000,
    load_in_4bit = True,
    token =userdata.get('HF_TOKEN')
)
def chatbot(query):
  messages = [
      {"from": "system", "value":
       """
      Environment: ipython. Tools: brave_search, wolfram_alpha. Cutting Knowledge Date: December 2023. Today Date: 4 October 2024\n
      You are a coding assistant with expert with everything\n
      Ensure any code you provide can be executed \n
      with all required imports and variables defined. List the imports.  Structure your answer with a description of the code solution. \n
      write only the code. do not print anything else.\n
      use ipython for search tool. \n
      debug code if error occurs. \n
      Here is the user question: {question}
      """
       },
      {"from": "human", "value": query},
  ]
  inputs = tokenizer.apply_chat_template(messages, tokenize = True, add_generation_prompt = True, return_tensors = "pt").to("cuda")

  text_streamer = TextStreamer(tokenizer)
  _ = model.generate(input_ids = inputs, streamer = text_streamer, max_new_tokens = 2048, use_cache = True)

Execute code (Make sure to use virtual environments)

python3 -m venv env
source env/bin/activate

Execution code responses from Llama

Please use execute python code function for local. For langchain, please use Python REPL() to execute code

execute code funciton locally in python:

def execute_Python_code(code):
     # A string stream to capture the outputs of exec
    output = io.StringIO() 
    try:
        # Redirect stdout to the StringIO object
        with contextlib.redirect_stdout(output):  
            # Allow imports 
            exec(code, globals())
    except Exception as e:
        # If an error occurs, capture it as part of the output
        print(f"Error: {e}", file=output)  
    return output.getvalue()

Langchain python Repl

• Install

!pip install langchain_experimental

Code:

from langchain_core.tools import Tool
from langchain_experimental.utilities import PythonREPL

python_repl = PythonREPL()

# You can create the tool to pass to an agent
repl_tool = Tool(
    name="python_repl",
    description="A Python shell. Use this to execute python commands. Input should be a valid python command. If you want to see the output of a value, you should print it out with `print(...)`.",
    func=python_repl.run,
)
repl_tool(outputs[0]["generated_text"][-1])

Safety inputs/ outputs procedures

Fo all inputs, please use Llama-Guard: meta-llama/Llama-Guard-3-8B for safety classification. Go to model card Llama-Guard

Uploaded model

• Developed by: EpistemeAI
• License: apache-2.0
• Finetuned from model : EpistemeAI/Fireball-Meta-Llama-3.1-8B-Instruct-Agent-0.003-128K-code

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	22.48
IFEval (0-Shot)	55.15
BBH (3-Shot)	26.74
MATH Lvl 5 (4-Shot)	12.01
GPQA (0-shot)	7.27
MuSR (0-shot)	6.79
MMLU-PRO (5-shot)	26.89