Your model should dream while you sleep. 😴
Today we’re open-sourcing Songe.
It’s a local training engine that turns your conversations, instructions, and corrections into training data, then fine-tunes your local model on your own tasks and preferences.
It all runs on your own machine, using compute that would otherwise sit idle.
Your data stays on your device. The model that gets better is yours.
pip install songe
https://t.co/QIHc4caigj
@vveerrgg Good point. Today each new version is checked against the old one before it's used, and you can roll back. Measuring drift over many dreams is a good idea.
@Connex01 Agreed, that's the hard part. Rehearsal plus checking that nothing got worse is where we are today. Not solved yet, we're testing a few newer ideas for exactly this.
@crislenta It does that already :) It doesn't train on raw chats, only on lessons from the moments you corrected it, and a new version is only used if it's better.
@pushkar_tbpl20 A newer correction becomes a new lesson and the next dream learns it. You can also remove a preference or roll back to an earlier version. Memory outside and weights inside probably go together.
@CeciNowBe Good question. A new version only gets used if it does better on your own conversations without breaking what already worked, and old skills get rehearsed over time. You can always roll back too.
@AyzeLYC@kawaidgaf Not yet, training runs on the same GPU as the model for now. Splitting it across two cards is a good idea though. Which cards do you have?
Thanks for all the love on the launch 😴
Songe update:
TL;DR: bigger models, MoE + multimodal, more control. Next: RL and frontier distillation.
- Bigger models on one GPU: CPU offload and a paged optimizer are in. Qwen3.8-27B now trains on a single 32 GB RTX 5090.
- Supporting MoE and multimodal models: Inkling-2-Mini (26B) trains with its experts kept packed, 16 GB instead of 49 GB.
- More control: go back to any earlier version, undo a Quick Dream, or remove a single preference from the dashboard.
Next:
- Reinforcement learning. Your model builds a sandbox from your kind of work, tries tasks there, and learns from what worked.
- Frontier distillation. Distill a frontier model into your local model, on your own tasks.
Ideas welcome :)
https://t.co/j5fVhObCgQ
Your model should dream while you sleep. 😴
Today we’re open-sourcing Songe.
It’s a local training engine that turns your conversations, instructions, and corrections into training data, then fine-tunes your local model on your own tasks and preferences.
It all runs on your own machine, using compute that would otherwise sit idle.
Your data stays on your device. The model that gets better is yours.
pip install songe
https://t.co/QIHc4caigj
@harshTalksAI On an Apple Silicon Mac, yes. On a PC with integrated graphics it will run and record, but training needs an NVIDIA or AMD card for now. What machine do you have?