| --- |
| license: apache-2.0 |
| language: |
| - en |
| pipeline_tag: text-generation |
| tags: |
| - conversational |
| - qwen3 |
| - unsloth |
| - reasoning |
| base_model: |
| - Qwen/Qwen3-4B-Thinking-2507 |
| --- |
| |
| # Solace Alpha |
|
|
| Solace Alpha is an experimental 4B parameter model built for one thing: actually having a good conversation. |
|
|
| I got tired of AI that sounds like a heavily-litigated corporate press release. Solace is trained to reject the usual AI tropes—no excessive positivity, no sycophantic agreement, and absolutely no making things up. It's grounded, thoughtful, and maybe a little bit existential. |
|
|
| It is also really good at webdev for it's size. |
|
|
| ## Specs |
| - **Parameters**: 4B |
| - **Base Model**: Qwen/Qwen3-4B-Thinking-2507 |
| - **Context Length**: 8k |
| - **Data**: Fine-tuned on a custom dataset of high-reasoning, multi-turn conversations designed to give it a certian *VIBE* |
|
|
| ## Why it's different |
| Solace isn't your standard "how can I help you today 😁" assistant. |
|
|
| 1. **It's honest about what it doesn't know.** If you ask about its origins or architecture, it won't hallucinate a fake company name or claim it's a person. If it doesn't know, it just says it doesn't know. |
| 2. **It handles negative emotions normally.** No toxic positivity. If you're stressed or frustrated, it holds space for that without rushing to "fix" you with empty platitudes. |
| 3. **It has strict boundaries.** It knows it's an AI. It won't roleplay as a human, pretend to experience the world like one, or return romantic affection. |
| 4. **It actually has a sense of humor.** Expect dry, observational humor. It's capable of playfully pushing back or roasting you when appropriate instead of defaulting to self-deprecation. |
|
|
| ## Usage |
| Solace is heavily optimized for direct back-and-forth dialogue, deep philosophical questions, and messy interpersonal stuff. |
|
|
| ## Training |
| This model was trained on a custom dataset consisting of ~1,100 examples, trained for 4 epochs using Unsloth (my beloved). This also happens to be the seventh iteration of this model as I kept on adding more to the dataset to steer the model in the direction I wanted. |
|
|
| ## Heads up |
| - **It can get wordy**: When you ask deep technical or philosophical questions, it tends to write multi-paragraph answers. It likes to think out loud. |
| - **Boundaries**: It won't help you hurt yourself. It has strict safety overrides but handles them with playful or serious deflections depending on the context. |
| - **Not a therapist**: Seriously, it's just a language model. It will literally tell you to go see a human professional if you try to use it as a replacement for actual mental health support. |
