Spaces:
Sleeping
Sleeping
| title: Outreach Message Generator | |
| sdk: gradio | |
| emoji: 🔥 | |
| colorFrom: indigo | |
| colorTo: blue | |
| short_description: This model is a fine-tuned version of google/gemma-2b traine | |
| # Gemma-2B Fine-Tuned for Outreach Message Generation | |
| ## Model Description | |
| This model is a fine-tuned version of **google/gemma-2b** trained to generate personalized LinkedIn outreach messages for business communication. The model takes contextual information about a prospect and generates a professional first-level outreach message suitable for networking or business development. | |
| The model was fine-tuned using **parameter-efficient fine-tuning (LoRA)** with the **PEFT library** and **Hugging Face Transformers**. To reduce memory usage during training, **4-bit quantization using BitsAndBytes** was applied. | |
| The model is designed to generate context-aware outreach messages based on structured inputs such as industry, company name, prospect designation, and background information. | |
| --- | |
| ## Intended Use | |
| The model is intended for generating personalized outreach messages for: | |
| - LinkedIn networking | |
| - Sales outreach | |
| - Business development communication | |
| - Professional introduction messages | |
| The model accepts structured contextual inputs and generates a professional outreach message based on the provided information. | |
| Example input fields include: | |
| - Prospect Name | |
| - Designation | |
| - Company | |
| - Industry | |
| - Prospect Background | |
| - Sender Background | |
| --- | |
| ## Prompt Format | |
| The model was trained using an instruction-style prompt-response format. | |
| Example prompt structure: | |
| PROMPT: | |
| Generate a personalized LinkedIn message for {Prospect Name}, | |
| {Designation} at {Company} ({Industry}) | |
| Prospect Background: | |
| {Prospect Background} | |
| My Background: | |
| {Sender Background} | |
| RESPONSE: | |
| The model generates a personalized outreach message as the response. | |
| --- | |
| ## Training Data | |
| The training dataset consisted of structured outreach communication examples containing: | |
| - Prospect Name | |
| - Designation | |
| - Company Name | |
| - Industry | |
| - Prospect Information | |
| - Sender Information | |
| - First-Level Outreach Message | |
| The dataset was formatted into **prompt–completion pairs** to train the model in an instruction-following manner. | |
| --- | |
| ## Training Procedure | |
| The model was fine-tuned using **Supervised Fine-Tuning (SFT)** with **LoRA adapters**. | |
| ### Key training configuration | |
| - **Base Model:** google/gemma-2b | |
| - **Fine-Tuning Method:** LoRA (PEFT) | |
| - **Quantization:** 4-bit NF4 (BitsAndBytes) | |
| - **Optimizer:** paged_adamw_8bit | |
| - **Epochs:** 5 | |
| - **Learning Rate:** 2e-4 | |
| - **Batch Size:** 2 | |
| - **Gradient Accumulation Steps:** 4 | |
| - **Mixed Precision Training:** FP16 | |
| LoRA adapters were applied to the transformer attention layers: | |
| - `q_proj` | |
| - `k_proj` | |
| - `v_proj` | |
| - `o_proj` | |
| This enabled efficient fine-tuning while keeping most of the base model weights frozen. | |
| --- | |
| ## Evaluation | |
| Basic evaluation was performed using generated responses compared against reference messages. **ROUGE metrics** were used to measure similarity between generated and reference messages. | |
| The model demonstrates the ability to generate structured outreach messages that follow the tone and style of professional LinkedIn communication. | |
| --- | |
| ## Example Usage | |
| Example Python inference code: | |
| ```python | |
| from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline | |
| model_name = "KirubaLS/gemma2b-merged1" | |
| tokenizer = AutoTokenizer.from_pretrained(model_name) | |
| model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto") | |
| generator = pipeline("text-generation", model=model, tokenizer=tokenizer) | |
| prompt = """ | |
| ### PROMPT: | |
| Generate a personalized LinkedIn message for John Doe, | |
| Senior Data Engineer at Stripe (FinTech) | |
| Prospect Background: | |
| Works on payment infrastructure and scaling distributed systems. | |
| My Background: | |
| AI engineer building data intelligence platforms. | |
| ### RESPONSE: | |
| """ | |
| result = generator(prompt, max_new_tokens=200) | |
| print(result[0]["generated_text"]) | |
| Limitations | |
| The model may produce generic outreach messages when limited contextual information is provided. Output quality improves when richer contextual information is included in the prompt. | |
| The model is intended for professional communication use cases and should be reviewed before sending messages in real-world scenarios. | |
| License | |
| This model follows the license of the base model google/gemma-2b. | |
| Acknowledgements | |
| This project was built using: | |
| Hugging Face Transformers | |
| PEFT (Parameter-Efficient Fine-Tuning) | |
| BitsAndBytes quantization | |
| PyTorch |