email-qwen3-0.6b/README.md

---
license: apache-2.0
language:
- en
tags:
- email
- cold-outreach
- text-generation
- qwen3
- fine-tuned
base_model: Qwen/Qwen3-0.6B
pipeline_tag: text-generation
---

# Email-Qwen3-0.6B — Fine-tuned for Email Generation

A fine-tuned Qwen3 0.6B model specialized in generating professional emails from simple prompts. Trained on 130k curated email examples with 5 rounds of rejection sampling alignment.

## Model Details

- **Base model:** Qwen/Qwen3-0.6B
- **Training:** SFT on 130k prompt-email pairs + 5 rounds rejection sampling fine-tuning
- **Quantized version:** Q4_K_M (378MB) available for local inference via llama.cpp
- **Use case:** Cold outreach, thank-you, request, apology, invitation, congratulations, and 10+ other email types

## Usage

### With llama.cpp (recommended)

```bash
# Download the GGUF quantized version
# Start the server
llama-server -m email-qwen3-06b-q4_k_m.gguf --host 127.0.0.1 --port 8081 -c 2048

# Generate an email
curl http://127.0.0.1:8081/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [
      {"role": "system", "content": "You are an email writing assistant. Write a polished email body for the given request."},
      {"role": "user", "content": "Cold outreach to the CTO at Stripe about our developer tools platform"}
    ],
    "max_tokens": 256,
    "temperature": 0.7
  }'
```

### With Transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("CharlieGreenman/email-qwen3-0.6b")
tokenizer = AutoTokenizer.from_pretrained("CharlieGreenman/email-qwen3-0.6b")

messages = [
    {"role": "system", "content": "You are an email writing assistant. Write a polished email body for the given request."},
    {"role": "user", "content": "Thank Sarah for helping with the presentation last week"},
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.7, do_sample=True)
print(tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True))
```

## Tips for Best Results

- **One paragraph at a time:** This model performs best when asked to generate individual paragraphs rather than full multi-paragraph emails. Generate each paragraph with a focused prompt, then assemble.
- **Keep prompts specific:** Include the recipient's name, company, role, and topic for better personalization.
- **Use best-of-N:** Generate 3-5 variants and pick the best one. Small models benefit significantly from selection.
- **Temperature 0.7-0.8** works well for email generation.

## Training Data

- 50,000 diverse email prompts across 17 email types
- 34,055 high-quality prompt-email pairs (scored 80+ by our quality scorer)
- 96,034 section-level examples (individual email paragraphs)
- 5 rounds of rejection sampling using best-of-5 selection with quality scoring

## Supported Email Types

Cold outreach, follow-up, newsletter, transactional, welcome, personal, request, meeting, FYI, thank-you, confirmation, apology, introduction, invitation, deadline, congratulations, and freeform.

## Limitations

- Best for common email types; may struggle with unusual or highly creative prompts
- Generates email body text; subject lines should be handled separately
- Small model (0.6B) — quality improves significantly with best-of-N selection and post-processing
- May occasionally hallucinate company names or statistics

## License

Apache 2.0