| --- |
| language: |
| - en |
| tags: |
| - text-generation-inference |
| - text-generation |
| --- |
| |
| <img src="https://cdn-uploads.huggingface.co/production/uploads/650707344a8839a8bd85ae2f/Usz0Okcz85HrkdGLpMLKu.png" style="height: 400px;"> |
|
|
| # Synth 2.5 Pro Preview |
|
|
|
|
| > [!Note] |
| > This is a preview model, it may deliver performance below that of the final model.<br> |
| > This preview is a model finetuned with purely **SFT**, RLAIF has **not** been applied yet.<br> |
| > Alongside this, creative performance is currently best on **non-thinking mode** for this preview.<br> |
|
|
| Synth 2.5 Pro is the next model in the Synth line-up, based on Gemma 4 26BA4B, it has 26B total parameters and 4B active parameters. |
|
|
| ### Compared to Synth 2 |
| Compared to Synth 2, 2.5 has the following changes/improvements: |
|
|
| - Trained on real-world creative data/usage from SOTA models (and past-sota, user-perferred models) |
|
|
| - Optimized for following user-wanted qualities (described [here](https://lucidity.sh/research/creative.html)) via RLAIF (READ NOTE FOR PREVIEW) |
|
|
| - Hybrid reasoning support, for optional further in-depth creative reasoning |
|
|
| ### Usage |
|
|
| #### Composite and LuciditySH Platform |
|
|
| You can test Synth 2.5 Pro at [Composite](https://composite.lucidity.sh/) (creative specific) or LuciditySH Platform (coming soon) for free, with up to 100 free requests a day for easy testing/usage. |
|
|
| #### Local |
|
|
| You can use Synth 2.5 Pro on Llama.cpp or VLLM |
|
|
| #### Generation Parameters |
|
|
| It is recommended to use Synth 2.5 Pro with the following generation settings: |
|
|
| - Temperature: 0.8-1 |
| - Top-P: 0.95 |
| - Top-K: 0 |
|
|
| ### Training |
|
|
| For the SFT stage of Synth 2.5's training, it was trained on a closed dataset comprised of real-world creative interactions with the following models, with the amount of interactions: |
|
|
| | Family | Samples | |
| |---|---:| |
| | Gemini (2.5 Pro) | 1 | |
| | Gemini (3.X Pro/3.7) | 1 | |
| | DeepSeek (V3 0324) | 5 | |
| | DeepSeek (R1 0528) | 8 | |
| | DeepSeek (V4 Pro) | 804 | |
| | GLM (5.X) | 1981 | |
| | GLM (4.X) | 58 | |
| | Kimi (k2.X) | 284 | |
| | Kimi (k3) | 3 | |
| | Minimax M3 | 69 | |
| | StepFun | 40 | |
| | **total** | **3254** | |
|
|
|
|
| <img src="https://cdn-uploads.huggingface.co/production/uploads/650707344a8839a8bd85ae2f/OMfjLL7rzSJQKOZIVL6PR.png" style="height: 350px;"> |
|
|
| For model replication, close data to our closed dataset can be found in our [PIPKIN datasets](https://huggingface.co/datasets/LucidityAI/PIPKIN-Creative-174k). |
|
|
| ### Limitations (for Preview) |
|
|
| As per the active parameter count of this model (4B) and the fact that it is only in preview, it comes with the following limitations, as flagged by Composite users: |
|
|
| - Synth 2.5 Pro Preview is unstable at higher temperatures (<1) |
| - Synth 2.5 Pro Preview may seem predictable, even at a higher temperature (0.8-1) |
|
|
| ### Considerations |
|
|
| Synth 2.5 Pro is capable of generating content that is harmful, illegal and/or generally NSFW, as with any LLM. It is recommended to put Synth 2.5 behind a moderation model or other safety layer for real world deployment. |
