AshankGupta's picture
Updated README.md File
227dac8 verified
metadata
base_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
library_name: llama.cpp
pipeline_tag: text-generation
tags:
  - tinyllama
  - email-reply
  - gguf
  - ollama
  - local-ai
license: mit
language:
  - en

TinyLlama Email Reply Generator

A fine-tuned TinyLlama model for generating professional email replies from incoming emails.

The adapter was trained on the Enron Email Reply Dataset to learn professional communication patterns.


Model Overview

  • Base Model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
  • Format: GGUF (Q4_K_M quantization)
  • Size: ~667 MB
  • Task: Email reply generation
  • Language: English

Quick Start with GGUF

Using Ollama (Recommended)

# Pull the model from Hugging Face
huggingface-cli download AshankGupta/tinyllama-email-reply tinyllama-chat.Q4_K_M.gguf --local-dir ./model

# Or download directly
curl -L -o model.gguf "https://huggingface.co/AshankGupta/tinyllama-email-reply/resolve/main/tinyllama-chat.Q4_K_M.gguf"

# Create Ollama model
ollama create tinyllama-email-reply -f Modelfile

# Run
ollama run tinyllama-email-reply

Using llama.cpp

# Download GGUF file from Hugging Face
curl -L -o tinyllama-email-reply.Q4_K_M.gguf \
  "https://huggingface.co/AshankGupta/tinyllama-email-reply/resolve/main/tinyllama-chat.Q4_K_M.gguf"

# Run inference
./llama.cpp/llama-cli -m tinyllama-email-reply.Q4_K_M.gguf -p "Write a professional email reply to: Can you send the invoice by tomorrow?"

Using Python

from llama_cpp import Llama

model = Llama(
    model_path="tinyllama-chat.Q4_K_M.gguf",
    n_ctx=1024,
)

prompt = """You are an AI assistant that writes professional email replies.

Email:
Can you send the invoice by tomorrow?

Reply:"""

output = model(prompt, max_tokens=120)
print(output["choices"][0]["text"])

Intended Use

This model is designed for:

  • Email reply suggestion systems
  • AI productivity tools
  • Email assistants
  • Local AI workflows
  • Research on small language models

Training Dataset

The model was trained using the Enron Email Reply Dataset, which contains real-world corporate email conversations.

Dataset characteristics:

  • ~15,000 email-reply pairs
  • Business and professional communication
  • Cleaned and formatted into instruction-style prompts

Training format example:

Instruction:
Generate a professional email reply.

Email:
Can you send the project report by tomorrow?

Reply:
Sure, I will send the report by tomorrow.

Training Details

  • Fine-tuning technique: LoRA
  • Training framework: Unsloth
  • Sequence length: 512 tokens
  • Optimizer: AdamW
  • Base architecture: TinyLlama 1.1B
  • Quantization: Q4_K_M

Example

Input Email

Hi,

Can you send the invoice by tomorrow?

Generated Reply

Sure, I will send the invoice by tomorrow.

Limitations

  • The model may produce generic replies.
  • Performance is limited by the small size of the base model.
  • It may occasionally generate repetitive outputs.
  • Not suitable for sensitive or confidential communications.

License

This model follows the license of the base model:

TinyLlama License

Please review the base model license before commercial usage.


Acknowledgements

  • TinyLlama team for the base model
  • Unsloth for efficient LoRA training
  • Enron Email Dataset for training data