mpt-7b-moe-nq-finetuned

MPT-7B-MoE Fine-tuned on Natural Questions

This model is a fine-tuned version of mosaicml/mpt-7b on the Natural Questions dataset.

Model Description

• Base Model: mosaicml/mpt-7b
• Architecture: Mixture-of-Experts (MoE) variant
• Training Dataset: Natural Questions (NQ) - annotated for expert routing
• Task: Question Answering / Text Generation

Training Details

The model was fine-tuned using:

• DeepSpeed with ZeRO optimization
• HuggingFace Accelerate
• Custom MoE expert annotations

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "vvijayk/mpt-7b-moe-nq-finetuned"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    trust_remote_code=True,
    torch_dtype="auto",
    device_map="auto"
)

# Example inference
question = "What is the capital of France?"
prompt = f"Question: {question}\nAnswer:"

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=128,
    temperature=0.7,
    do_sample=True,
    top_p=0.9,
)

answer = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(answer)

Training Configuration

Training was performed using the configuration in the repository. See train_mpt7b_moe_accelerate.py for details.

Limitations and Biases

This model inherits limitations and biases from the base MPT-7B model and the Natural Questions dataset. Users should be aware of potential biases in question-answering outputs.

Citation

If you use this model, please cite:

@misc{mpt-7b-moe-nq,
  author = {Your Name},
  title = {MPT-7B-MoE Fine-tuned on Natural Questions},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/vvijayk/mpt-7b-moe-nq-finetuned}}
}

Acknowledgements

• Base model by MosaicML
• Natural Questions dataset by Google Research
• Training infrastructure using DeepSpeed and HuggingFace Accelerate