Muigai1/productivity-assistant-qwen3

✨ Daily AI Productivity Assistant

A fine-tuned Qwen3-8B model specialized for everyday productivity tasks: email drafting, summarization, message rewriting, meeting notes, social media posts, and Q&A.

Model Details

• Base Model: Qwen/Qwen3-8B (Apache 2.0)
• Architecture: Qwen3, 8.2B parameters
• Context Length: 32K tokens
• Training Data: 510K high-quality conversations
  • trl-lib/Capybara (~10K)
  • HuggingFaceH4/ultrachat_200k (~200K)
  • Magpie-Align/Magpie-Pro-300K-Filtered (~300K)
• Training Method: SFT (Supervised Fine-Tuning)
• License: Apache 2.0 (fully monetizable)

Training Recipe

Based on published research:

• Magpie paper (arXiv:2406.08464): LR 2e-5, 2 epochs, effective batch 32
• Granite SFT guide (arXiv:2412.13337): constant scheduler, no warmup, packing=True

Parameter	Value
Learning Rate	2e-5
Epochs	2
Batch Size (per device)	1
Gradient Accumulation	32
Effective Batch Size	32
Max Sequence Length	4096
Scheduler	constant
Warmup	0
Packing	True
Precision	bfloat16

Usage

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "Muigai1/productivity-assistant-qwen3",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(
    "Muigai1/productivity-assistant-qwen3",
    trust_remote_code=True,
)

messages = [{"role": "user", "content": "Draft a professional email about a project delay."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7, top_p=0.8)
response = tokenizer.decode(outputs[0][len(inputs.input_ids[0]):], skip_special_tokens=True)
print(response)

Response Modes

• Fast mode (enable_thinking=False): Quick replies for emails, summaries, rewrites
• Deep Think mode (enable_thinking=True): Complex analysis, creative writing, coding

Business Model

Product: "Daily AI Productivity Assistant"

Target Market: Professionals, freelancers, students, small business owners who spend 2+ hours daily on writing and communication.

Pricing Tiers:

• Free: 50 messages/month, Fast mode only
• Pro ($9.99/month): Unlimited messages, Deep Think mode, email templates, priority speed
• Team ($29.99/user/month): Shared workspace, brand voice training, analytics, API access

Revenue Projections:

• 1,000 Pro users = $9,990/month
• 100 Team users (5 seats avg) = $14,995/month
• Total at 1,100 customers: ~$25K/month

Distribution

1. Hugging Face Space (free hosting with HF Pro)
2. Chrome Extension (Gmail, LinkedIn, Slack integration)
3. API (per-token pricing for developers)
4. Mobile App (iOS/Android subscription)

Competitive Advantage

• Trained on 510K diverse conversations vs. generic instruction tuning
• Dual-mode responses (Fast/Deep Think) from Qwen3 architecture
• Apache 2.0 license = no commercial restrictions
• 119 language support for global expansion

Files

• train.py — Complete fine-tuning script
• app.py — Gradio demo for deployment
• requirements.txt — Dependencies

Citation

@misc{qwen3,
  title={Qwen3 Technical Report},
  author={Qwen Team},
  year={2025},
  eprint={2505.09388},
  archivePrefix={arXiv},
}

@misc{magpie,
  title={Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing},
  author={Xu, Zhangchen et al.},
  year={2024},
  eprint={2406.08464},
  archivePrefix={arXiv},
}