| --- |
| 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 |
| --- |
| |
| <p align="center"> |
| <img src="https://huggingface.co/spaces/Lythri/README/resolve/main/Banner_flat.png" width="600" alt="Lythri"/> |
| </p> |
|
|
| <p align="center"> |
| <a href="https://huggingface.co/Lythri">Hugging Face</a> | |
| <a href="https://www.modelscope.cn/profile/Spike8086">ModelScope</a> | |
| <a href="https://zenodo.org/records/23179312">Technical Report</a> |
| </p> |
|
|
| ## Introduction |
|
|
| **Lythri-Tada-4B** is part of the **Lythri 2.0** family. It answers directly without a thinking phase, while reinforce model's emotional understanding. |
|
|
| Compared with the previous-generation Lythri-4B-A2B, Tada brings three main upgrades: |
|
|
| - **Stronger instruction following.** Tada scores 39.19 on IFEval (prompt-level strict), **surpassing even the Lythri-7B-A4B (31.05)**. |
| - **Better multi-turn conversations.** Tada keeps track of earlier context and user requests more consistently over long chats, and carries the emotional thread from turn to turn. |
| - **System prompt support.** Unlike Lythri-4B-A2B, Tada follows system prompts, so you can set a persona, tone or rules for your app directly in the system message. |
|
|
| Long-tail knowledge, such as graduate-level science questions, is slightly weaker than Lythri-4B-A2B, while overall performance improves **substantially**. |
|
|
| ## Writing Naturalness |
|
|
| <p align="center"> |
| <img src="https://huggingface.co/spaces/Lythri/README/resolve/main/Tada_Slop.png" width="800" alt="EQ-Bench Slop Score"/> |
| </p> |
|
|
| On the [EQ-Bench Slop Score](https://eqbench.com/slop-score.html), which measures how much a model relies on typical AI phrasing (lower is better), **Lythri-Tada-4B scores 39.0 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 |
|
|
| <table width="100%" style="display: table; width: 100%; border-collapse: collapse; text-align: center;"> |
| <thead> |
| <tr style="border-bottom: 2px solid #338f65;"> |
| <th width="40%" style="text-align: left; padding: 10px 16px;">Benchmark</th> |
| <th width="30%" style="color: #338f65; white-space: nowrap; padding: 10px 16px;">Lythri-4B-A2B</th> |
| <th width="30%" style="background: rgba(51,143,101,0.12); color: #338f65; white-space: nowrap; padding: 10px 16px;">Lythri-Tada-4B</th> |
| </tr> |
| </thead> |
| <tbody> |
| <tr style="background: rgba(51,143,101,0.06);"> |
| <td colspan="3" style="text-align: left; color: #338f65; font-weight: bold;">Knowledge</td> |
| </tr> |
| <tr><td style="text-align: left;">MMLU</td><td>55.23</td><td style="background: rgba(51,143,101,0.12); color: #338f65; font-weight: bold;">55.41</td></tr> |
| <tr><td style="text-align: left;">ARC-C</td><td>52.99</td><td style="background: rgba(51,143,101,0.12); color: #338f65; font-weight: bold;">56.57</td></tr> |
| <tr style="background: rgba(51,143,101,0.06);"> |
| <td colspan="3" style="text-align: left; color: #338f65; font-weight: bold;">Reasoning</td> |
| </tr> |
| <tr><td style="text-align: left;">HellaSwag</td><td>72.80</td><td style="background: rgba(51,143,101,0.12); color: #338f65; font-weight: bold;">74.23</td></tr> |
| <tr style="background: rgba(51,143,101,0.06);"> |
| <td colspan="3" style="text-align: left; color: #338f65; font-weight: bold;">General</td> |
| </tr> |
| <tr><td style="text-align: left;">SocialIQA</td><td>49.80</td><td style="background: rgba(51,143,101,0.12); color: #338f65; font-weight: bold;">51.43</td></tr> |
| <tr style="background: rgba(51,143,101,0.06);"> |
| <td colspan="3" style="text-align: left; color: #338f65; font-weight: bold;">Science</td> |
| </tr> |
| <tr><td style="text-align: left;">GPQA Diamond</td><td>28.79</td><td style="background: rgba(51,143,101,0.12); color: #338f65; font-weight: bold;">21.21</td></tr> |
| <tr style="background: rgba(51,143,101,0.06);"> |
| <td colspan="3" style="text-align: left; color: #338f65; font-weight: bold;">Math</td> |
| </tr> |
| <tr><td style="text-align: left;">GSM8K</td><td>28.81</td><td style="background: rgba(51,143,101,0.12); color: #338f65; font-weight: bold;">27.45</td></tr> |
| <tr style="background: rgba(51,143,101,0.06);"> |
| <td colspan="3" style="text-align: left; color: #338f65; font-weight: bold;">Code / Instruction</td> |
| </tr> |
| <tr><td style="text-align: left;">HumanEval</td><td>28.66</td><td style="background: rgba(51,143,101,0.12); color: #338f65; font-weight: bold;">24.39</td></tr> |
| <tr><td style="text-align: left;">IFEval</td><td>26.43</td><td style="background: rgba(51,143,101,0.12); color: #338f65; font-weight: bold;">39.19</td></tr> |
| </tbody> |
| </table> |
| |
| All benchmarks are evaluated with their official standard settings and the chat template applied, reflecting real-world inference conditions. Lythri-4B-A2B is evaluated with thinking; Lythri-Tada-4B answers directly without thinking. |
|
|
| ## Quickstart |
| ```python |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| import torch |
| |
| model_path = "Lythri/Lythri-Tada-4B" |
| 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-Tada-4B supports system prompts, which Lythri-4B-A2B does not. 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-Tada-4B starts from the Gemma4-E2B PT and training on SFT with CoT, followed by instant-reply SFT with multi-turn companion conversations, and online preference optimization on only writing with responses judged by Gemma2-27B. |
|
|
| ## 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!** |
