---
license: apache-2.0
base_model: google/gemma-4-E2B
tags:
- emotional-support
- companion
- on-device
- conversational
- emotion
- doi:10.5281/zenodo.23179311
library_name: transformers
pipeline_tag: text-generation
---
Hugging Face |
ModelScope |
Technical Report
## Introduction
**Lythri-Wink-2B** is part of the **Lythri 2.0** family, with **2.52B total parameters** and **1.48B activated parameters**. It is a compact, direct-response model optimized for on-device conversation and emotional support.
Compared with the larger Lythri-Tada-4B, Wink focuses on three practical features:
- **Smaller deployment footprint.** Wink substantially reduces both total and activated parameters.
- **Multi-turn conversations.** Wink is trained on multi-turn companion conversations and earlier-turn context.
- **System prompt support.** Like Lythri-Tada-4B, Wink follows system prompts, so you can set a persona, tone or rules for your app directly in the system message.
Wink trades knowledge, reasoning, math and coding performance for its smaller deployment footprint; it is not intended to match the larger Lythri-Tada-4B on general benchmarks.
## Writing Naturalness
Within the same family, **Lythri-Tada-4B scores 39.0 on the [EQ-Bench Slop Score](https://eqbench.com/slop-score.html)** and writes more naturally than GPT-4o (47.7) and DeepSeek-R1 (67.9), far ahead of Qwen3-4B (62.2) and Gemma-3-4B (85.2).
## General Benchmarks
| Benchmark |
Lythri-Tada-4B |
Lythri-Wink-2B |
| Knowledge |
| MMLU | 55.41 | 39.32 |
| ARC-C | 56.57 | 41.13 |
| Reasoning |
| HellaSwag | 74.23 | 58.65 |
| General |
| SocialIQA | 51.43 | 45.24 |
| Science |
| GPQA Diamond | 21.21 | 23.23 |
| Math |
| GSM8K | 27.45 | 13.04 |
| Code / Instruction |
| HumanEval | 24.39 | 7.93 |
| IFEval | 39.19 | 23.48 |
The listed benchmarks use their standard task settings and few-shot counts with the chat template applied. Lythri-Wink-2B is evaluated as a Q8_0 GGUF.
## Quickstart
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_path = "Lythri/Lythri-Wink-2B"
tok = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(
model_path, dtype=torch.bfloat16, device_map="auto"
)
messages = [{"role": "user", "content": "My friend just lost their job and seems really down. What should I say to them?"}]
chat = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(chat, return_tensors="pt", add_special_tokens=False).to(model.device)
with torch.inference_mode():
out = model.generate(
**inputs,
max_new_tokens=1024,
do_sample=True,
temperature=0.8,
top_p=0.95,
eos_token_id=[1, 106],
)
print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
```
### Recommended Generation Config
```python
generation_config = {
"temperature": 0.8,
"top_p": 0.95,
"top_k": 64,
"max_new_tokens": 1024,
"repetition_penalty": 1.05,
"do_sample": True,
"eos_token_id": [1, 106],
}
out = model.generate(**inputs, **generation_config)
```
### System Prompt (Optional)
Lythri-Wink-2B supports system prompts. It works well without one; to set a persona or rules for your app, add a system message:
```python
messages = [
{"role": "system", "content": "You are Lythri, a warm and caring companion. Reply in the same language as the user."},
{"role": "user", "content": "I had a really rough day at work."},
]
```
## Compute
Lythri-Wink-2B is structurally sheared from the Gemma4-E2B PT model and recovered through teacher distillation. It is then trained with CoT SFT, followed by instant-reply SFT with multi-turn companion conversations and a small writing-only online preference-optimization stage mainly judged by Gemma2-27B,and Gemma4-E4B as a language judge model for language consistency.
## Limitations
- Lythri is optimized for conversation and emotional understanding, not for math, coding or complex reasoning.
- Lythri is not a substitute for professional mental health support. If you or someone you know is in crisis, please contact local emergency services or a crisis helpline.
- Like all language models, it can produce inaccurate or inappropriate content.
## License
Lythri is built on Gemma 4 and is released under the Apache License 2.0.
## Citation
```bibtex
@techreport{li2026lythri,
title = {Lythri Technical Report},
author = {Li, Jiawen},
year = {2026},
institution = {Zenodo},
doi = {10.5281/zenodo.23179311},
url = {https://doi.org/10.5281/zenodo.23179311}
}
```
## Support
**The whole training process is self-funded. If you like our model, please click a free like — that means a lot to me as a student!**