---
license: apache-2.0
base_model: google/gemma-3-270m-it
tags:
  - sales
  - coaching
  - gemma
  - fine-tuned
  - conversational
language:
  - en
pipeline_tag: text-generation
---

# Gemma 3 Sales Coach 🎯

Fine-tuned **Gemma 3 270M IT** model for sales conversation coaching.

## Model Description

This model analyzes sales conversations and provides detailed coaching feedback, identifying:
- **Sales rep mistakes** with specific turn numbers
- **Why each mistake is problematic** 
- **Better alternative responses**
- **Impact on conversion probability**

## Training Data

- **Dataset**: 99,056 synthetic sales conversations with coaching annotations
- **Industries**: 60+ industries covered
- **Personas**: 20+ buyer personas
- **Outcomes**: Lost deals, stalled deals, weak commitments

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "convaiinnovations/gemma3-sales-coach",
    torch_dtype=torch.float16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("convaiinnovations/gemma3-sales-coach")

# Format your conversation as instruction
conversation = '''[Turn 1] Sales Rep: Hi Sarah, this is Mike from CloudTech Solutions. How are you today?
[Turn 2] Prospect: I'm busy. What's this about?
[Turn 3] Sales Rep: Great! I wanted to tell you about our amazing cloud platform. It's the best in the market!
[Turn 4] Prospect: We already have AWS. Why would we switch?
[Turn 5] Sales Rep: Well, our platform is much better than AWS in every way.
[Turn 6] Prospect: That's a bold claim. Can you be more specific?
[Turn 7] Sales Rep: We have better uptime and support. Anyway, pricing starts at $10,000 per month.
[Turn 8] Prospect: That's expensive. What's included?
[Turn 9] Sales Rep: Everything you need. Trust me, it's worth it.
[Turn 10] Prospect: I'd need to see some case studies or ROI data.
[Turn 11] Sales Rep: I can send those later. So, are you ready to schedule a demo?
[Turn 12] Prospect: I don't think so. Send me some materials and I'll review when I have time.

Where did I make mistakes and how can I improve?'''

# Apply chat template (instruction format)
messages = [{"role": "user", "content": conversation}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.7,
    do_sample=True
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Training Details

- **Base Model**: google/gemma-3-270m-it
- **Method**: LoRA fine-tuning
- **LoRA Rank**: 32
- **Epochs**: 5
- **Hardware**: 2x NVIDIA T4 GPUs

## Limitations

- Optimized for B2B sales conversations
- Best results with 10+ turn conversations
- English language only

## License

Apache 2.0 (same as base Gemma model)