metadata

base_model: EpistemeAI/Fireball-Mistral-Nemo-Base-2407-sft-v2.2a
language:
  - en
license: apache-2.0
tags:
  - text-generation-inference
  - transformers
  - unsloth
  - mistral
  - trl
datasets:
  - candenizkocak/code-alpaca-297k
  - yahma/alpaca-cleaned
  - reciperesearch/dolphin-sft-v0.1-preference
pipeline_tag: text-generation
model-index:
  - name: Fireball-12B
    results:
      - task:
          type: text-generation
        dataset:
          name: dolphin-sft-v0.1-preference
          type: reciperesearch/dolphin-sft-v0.1-preference
        metrics:
          - name: MMLU_PRO
            type: MMLU
            value: 26.04
          - name: bbh
            type: bbh
            value: 30.67
          - name: IFEval
            type: IFeval
            value: 18.34
        source:
          name: Open LLM Leaderboard
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard

Fireball-12B

This model is super fine-tune to provide better coding and better response(from first fine-tune) than Llama-3.1-8B and Google Gemma 2 9B. Further fine tuned with ORPO method with dataset. Best use in alpaca(see Prompt instructions - Alpaca style prompt(recommended) ) instruct mode for best response, instead of chat mode.

reciperesearch/dolphin-sft-v0.1-preference

Benchmark

## Training Dataset Supervised fine-tuning with dataset: - candenizkocak/code-alpaca-297k - yahma/alpaca-cleaned

Model Card for Fireball-12B

The Heavy fine-tuned Mistral-Nemo-Base-2407 Large Language Model (LLM) is a pretrained generative text model of 12B parameters trained jointly by Mistral AI and NVIDIA, it significantly outperforms existing models smaller or similar in size.

For more details about this model please refer to our release blog post.

Key features

Released under the Apache 2 License
Pre-trained and instructed versions
Trained with a 128k context window
Trained on a large proportion of multilingual and code data
Drop-in replacement of Mistral 7B

Model Architecture

Mistral Nemo is a transformer model, with the following architecture choices:

Layers: 40
Dim: 5,120
Head dim: 128
Hidden dim: 14,436
Activation Function: SwiGLU
Number of heads: 32
Number of kv-heads: 8 (GQA)
Vocabulary size: 2**17 ~= 128k
Rotary embeddings (theta = 1M)

Guardrail/Moderation guide:

For guardrailing and moderating prompts against indirect/direct prompt injections and jailbreaking, please follow the SentinelShield AI GitHub repository: SentinelShield AI

Prompt Template: Alpaca (recommended)

plesee use Alpaca prompt

f"""Below is an instruction that describes a task. \
    Write a response that appropriately completes the request.

    ### Instruction:
    {x['instruction']}

    ### Input:
    {x['input']}

    ### Response:
    """

Demo

After installing mistral_inference, a mistral-demo CLI command should be available in your environment.

Prompt instructions - Alpaca style prompt(recommended):

f"""Below is an instruction that describes a task. \
    Write a response that appropriately completes the request.

    ### Instruction:
    {x['instruction']}

    ### Input:
    {x['input']}

    ### Response:
    """

Transformers

NOTE: Until a new release has been made, you need to install transformers from source:
pip install mistral_inference
pip install mistral-demo
pip install git+https://github.com/huggingface/transformers.git

If you want to use Hugging Face transformers to generate text, you can do something like this.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "EpistemeAI/Fireball-12B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)
inputs = tokenizer("Hello my name is", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=20)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Accelerator mode:

pip install accelerate #GPU A100/L4

from transformers import AutoModelForCausalLM, AutoTokenizer
from accelerate import Accelerator

# Initialize the accelerator
accelerator = Accelerator()

# Define the model ID
model_id = "EpistemeAI/Fireball-12B"

# Load the tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Load the model and prepare it for distributed setup using accelerate
model = AutoModelForCausalLM.from_pretrained(model_id)

# Move the model to the appropriate device using accelerate
model, = accelerator.prepare(model)

# Prepare inputs
inputs = tokenizer("Hello my name is", return_tensors="pt").to(accelerator.device)

# Generate outputs with the model
outputs = model.generate(**inputs, max_new_tokens=20)

# Decode and print the outputs
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Unlike previous Mistral models, Mistral Nemo requires smaller temperatures. We recommend to use a temperature of 0.3.

Note

EpistemeAI/Fireball-12B is a pretrained base model and therefore does not have any moderation mechanisms. Go to Guardrail/Moderation guide section for moderation guide

Citation for yahma/alpaca-cleaned dataset

@misc{alpaca,
  author = {Rohan Taori and Ishaan Gulrajani and Tianyi Zhang and Yann Dubois and Xuechen Li and Carlos Guestrin and Percy Liang and Tatsunori B. Hashimoto },
  title = {Stanford Alpaca: An Instruction-following LLaMA model},
  year = {2023},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/tatsu-lab/stanford_alpaca}},
}

Uploaded model

Developed by: EpistemeAI
License: apache-2.0
Finetuned from model : EpistemeAI/Fireball-Mistral-Nemo-Base-2407-sft-v2.2a

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