Bio2Tags-Lite/README.md

---
license: apache-2.0
language:
- en
tags:
- bio-to-tags
- tag-generation
- smollm2
- text-generation
- personality
- interests
- spiceechat
pipeline_tag: text-generation
library_name: transformers
---

<p align="center">
  <img src="https://huggingface.co/SpiceeChat/Bio2Tags-Qwen3.5-4B-SFT/resolve/main/Spiceechat.png" 
       alt="SpiceeChat" 
       width="1100" 
       height="1000" 
       style="border-radius: 50%; object-fit: cover;">
</p>

<p align="center">
  <a href="https://huggingface.co/HuggingFaceTB/SmolLM2-360M-Instruct"><img src="https://img.shields.io/badge/SmolLM2-360M-blue?logo=huggingface" alt="SmolLM2"></a>
  <a href="https://github.com/unslothai/unsloth"><img src="https://img.shields.io/badge/Fine‑Tuned-QLoRA-green" alt="QLoRA"></a>
  <a href="https://huggingface.co/SpiceeChat"><img src="https://img.shields.io/badge/SpiceeChat-🔥-orange" alt="SpiceeChat"></a>
  <a href="https://www.apache.org/licenses/LICENSE-2.0"><img src="https://img.shields.io/badge/License-Apache%202.0-yellow" alt="License"></a>
</p>

---

# 🏷️ Bio2Tags-Lite

**Because reading between the lines shouldn't require a psychology degree.**

Bio2Tags-Lite is a fine-tuned SmolLM2-360M model that reads personal biographies and returns clean, structured personality tags. Feed it a dating bio, a LinkedIn summary, or whatever someone wrote about themselves at 2am — it'll tell you what kind of person they actually are.

No rambling. No fluff. Just tags.

---

## ✨ Features

- **Lightweight**: 360M parameters — runs on hardware that would make a gamer cry
- **Fast**: Inference in milliseconds, because nobody has time to wait
- **Structured Output**: Clean comma-separated tags, every time
- **Plug & Play**: Works with Transformers out of the box, no PhD required
- **SpiceeChat Pipeline**: Pairs with Cinder-1.5B like peanut butter and heartbreak

---

## 🧪 Example

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "SpiceeChat/Bio2Tags-Lite",
    torch_dtype="auto",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("SpiceeChat/Bio2Tags-Lite")

def get_tags(bio):
    prompt = f"Extract personality tags from the bio below. Output ONLY comma-separated tags, nothing else.\n\nBio: {bio}\n\nTags:"
    messages = [{"role": "user", "content": prompt}]
    formatted = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer(formatted, return_tensors="pt").to(model.device)
    outputs = model.generate(**inputs, max_new_tokens=50, temperature=0.7, do_sample=True)
    return tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True).strip()

# Try it
print(get_tags("I love hiking at dawn, painting watercolors, and deep conversations about philosophy."))
# Output: nature-lover, artist, intellectual, deep-thinker
```

---

## 📊 Sample Outputs

| Bio | Tags |
|-----|------|
| "I'm a software engineer who loves late-night coding and playing jazz piano." | tech-savvy, creative, night-owl, music-enthusiast, artistic |
| "I spend my weekends trail running and evenings reading classic literature." | adventurous, nature-lover, bookworm, intellectual, quiet |
| "I'm a retired teacher who gardens, reads history books, and bakes sourdough." | intellectual, family-oriented, gardener, history-buff, old-soul |
| "As a digital nomad, my office changes weekly — from Bali cafes to Alpine cabins." | adventurous, creative, digital-nomad, spontaneous, tech-savvy |

*(Yes, the sourdough one is a stereotype. Yes, it's also always accurate.)*

---

## 📦 Installation

```bash
pip install transformers torch accelerate
```

That's it. No ritual sacrifices, no config files, no Stack Overflow rabbit holes.

---

## 🎯 Use Cases

- **Dating Apps**: Tag user bios automatically for smarter matching — because "I like long walks on the beach" means something very different than "I like long walks on the beach at 3am alone"
- **Social Media**: Generate relevant hashtags from profile descriptions
- **Recommender Systems**: Build personality-based recommendation engines
- **Content Analysis**: Extract structured metadata from unstructured text
- **SpiceeChat Pipeline**: Feed extracted tags into Cinder-1.5B for personalized compatibility advice

---

## 🛠️ Technical Details

| Detail | Value |
|--------|-------|
| **Base Model** | [SmolLM2-360M-Instruct](https://huggingface.co/HuggingFaceTB/SmolLM2-360M-Instruct) |
| **Fine-tuning Method** | QLoRA (4-bit quantization, rank-16 adapters) |
| **Training Framework** | Unsloth |
| **Training Data** | 1,387 hand-crafted (bio, tags) pairs |
| **Epochs** | 3 |
| **Learning Rate** | 1e-4 |
| **Sequence Length** | 512 tokens |
| **Hardware Used** | Google Colab T4 (free tier — yes, really) |
| **Final Size** | 724 MB (FP16) |
| **Min VRAM Required** | ~1.5 GB |

---

## ⚠️ Limitations

- **English only**: Other languages may produce results ranging from "creative" to "confidently wrong"
- **Training data size**: 1,387 examples is a solid start — more data is always on the roadmap
- **Tag granularity**: Captures the salient stuff, not every quirk (the model can't detect if someone is secretly obsessed with true crime podcasts)
- **Edge cases**: Very short bios, emoji-heavy text, or deeply abstract descriptions may surprise you

---

## 🧠 Part of the SpiceeChat Ecosystem

Bio2Tags-Lite is a core component of the SpiceeChat AI pipeline:

- 🏷️ **Bio2Tags-Lite** → Extracts personality tags from bios
- 🔥 **[Cinder-1.5B](https://huggingface.co/SpiceeChat/Cinder-1.5B)** → Personalized dating advice powered by those tags
- 🌐 **[dating-fatigue.com](https://dating-fatigue.com)** → Live tools for real humans trying to find real love

---

## 📜 License

Apache 2.0 — use it, modify it, ship it. Just give SpiceeChat a nod.

---

<div align="center">
  <sub>Built with ❤️ by <b>SpiceeChat</b></sub>
  <br>
  <sub>🔗 <a href="https://huggingface.co/SpiceeChat">huggingface.co/SpiceeChat</a></sub>
</div>