arunb74/llama-3.2-3b-nhm-production

Llama 3.2 3B NHM Production

A PEFT/LoRA fine-tuned adapter based on Meta Llama 3.2 3B Instruct.

This model is intended for multilingual NLP tasks related to the NHM (National Health Mission) use case.

Important: This repository contains the LoRA/PEFT adapter, not the complete base model weights. The base model meta-llama/Llama-3.2-3B-Instruct is required to use this adapter.

Model Details

• —Base model: meta-llama/Llama-3.2-3B-Instruct
• —Model type: PEFT / LoRA adapter
• —Architecture: Llama 3.2 3B Instruct
• —Fine-tuning method: LoRA
• —Framework: Hugging Face Transformers + PEFT
• —Task: Multilingual text generation / instruction following
• —Repository type: Fine-tuned adapter
• —Status: Production adapter

Repository Contents

This repository contains:

adapter_config.json
adapter_model.safetensors
chat_template.jinja
special_tokens_map.json
tokenizer_config.json
tokenizer.json
README.md

The large original Llama model weights are not included in this repository.

Base Model

The adapter was fine-tuned from:

meta-llama/Llama-3.2-3B-Instruct

The base model is gated by Meta on Hugging Face. Users must have the required access to the base model before loading this adapter.

Usage

Requirements

Install the required Python packages:

pip install torch transformers peft accelerate

Load the Adapter

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model_id = "meta-llama/Llama-3.2-3B-Instruct"
adapter_id = "arunb74/llama32-3b-nhm-production"

tokenizer = AutoTokenizer.from_pretrained(
    base_model_id
)

base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

model = PeftModel.from_pretrained(
    base_model,
    adapter_id
)

model.eval()

Generate Text

messages = [
    {
        "role": "user",
        "content": "Your question here"
    }
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

with torch.no_grad():
    outputs = model.generate(
        inputs,
        max_new_tokens=256,
        temperature=0.7,
        do_sample=True
    )

response = tokenizer.decode(
    outputs[0][inputs.shape[-1]:],
    skip_special_tokens=True
)

print(response)

Multilingual Usage

This adapter is designed for multilingual use cases. The actual supported languages and performance depend on the training data and fine-tuning configuration.

Example:

messages = [
    {
        "role": "user",
        "content": "Enter your multilingual NHM-related question here."
    }
]

Use the same language as the input when multilingual response behavior is desired.

Fine-Tuning

This model was fine-tuned using the LoRA (Low-Rank Adaptation) approach through Hugging Face PEFT.

Only the adapter parameters are stored in this repository.

The resulting adapter is approximately:

adapter_model.safetensors
~18 MB

This is significantly smaller than the original Llama 3.2 3B model because the full base model weights are not duplicated.

Adapter Configuration

The adapter configuration is stored in:

adapter_config.json

The adapter should be loaded on top of the corresponding base model:

meta-llama/Llama-3.2-3B-Instruct

Do not load the adapter as if it were a standalone causal language model.

Training Data

The adapter was fine-tuned for an NHM-related multilingual use case.

Specific training dataset details, including dataset size, language distribution, preprocessing, and licensing, should be documented separately if applicable.

Evaluation

Evaluation should be performed against the intended NHM multilingual production tasks.

Recommended evaluation dimensions include:

• —Response correctness
• —Instruction following
• —Multilingual performance
• —Factual consistency
• —Relevance
• —Safety
• —Hallucination rate
• —Response quality
• —Performance across supported languages

No standardized benchmark results are currently provided in this model card.

Limitations

This model inherits limitations from the underlying Llama 3.2 3B Instruct model and may also introduce additional limitations from the fine-tuning data.

Potential limitations include:

• —Hallucinated information
• —Incorrect or incomplete answers
• —Uneven performance across languages
• —Sensitivity to prompt wording
• —Domain-specific errors
• —Potential biases present in the training data
• —Reduced performance on tasks outside the fine-tuning domain

The model should not be assumed to provide authoritative medical, legal, financial, or other high-stakes advice.

Intended Use

This model is intended for:

• —NHM-related multilingual NLP applications
• —Research and development
• —Internal testing
• —Prototyping
• —Domain-specific text generation
• —Multilingual instruction-following applications

For production applications, responses should be appropriately validated and monitored.

Out-of-Scope Use

This model should not be used as the sole decision-making system for:

• —Medical diagnosis
• —Medical treatment decisions
• —Emergency medical decisions
• —Legal decisions
• —Financial decisions
• —Any other high-stakes automated decision

Human review and appropriate domain-specific safeguards should be used where necessary.

Hardware and Software

Base Model

Meta Llama 3.2 3B Instruct

Fine-Tuning Framework

Hugging Face PEFT

PEFT Version

0.10.0

The exact hardware, training duration, batch size, learning rate, number of epochs, and other training parameters should be added here if available.

License

The base model is subject to the license and usage terms provided by Meta for Llama 3.2.

Users should review and comply with the applicable Llama 3.2 license and acceptable-use requirements before using the model.

This adapter repository does not replace the licensing requirements of the base model.

Citation

If you use this adapter in research or an application, please cite the underlying Llama model and the relevant project or dataset documentation.

Acknowledgements

This model is based on:

Meta Llama 3.2 3B Instruct

and uses the Hugging Face PEFT framework for parameter-efficient fine-tuning.

Contact

For questions regarding this adapter or the associated NHM multilingual project, please refer to the project documentation and repository maintainers.