Great work!
Hey Jackrong,
Really impressive work on v3. The shift from pure distillation to structural alignment is smart, and the results back it up.
I've been experimenting with Qwen3.5 for study purposes and your fine-tune stood out. Would love to dig deeper into the training side of things — not the data (totally get keeping that private), but more the recipe itself.
No pressure at all, just curious. Either way — excellent work, keep it up!!!
Hey Jackrong,
Really impressive work on v3. The shift from pure distillation to structural alignment is smart, and the results back it up.
I've been experimenting with Qwen3.5 for study purposes and your fine-tune stood out. Would love to dig deeper into the training side of things — not the data (totally get keeping that private), but more the recipe itself.
No pressure at all, just curious. Either way — excellent work, keep it up!!!
Hey, really appreciate the support — it means a lot!
I’ll try to organize and share more details when I have some time. To be honest, everything I’ve done is based on the Unsloth repo and their guidance.
One thing I’ve noticed is that the quality of publicly available distillation datasets varies a lot. Some of them are honestly quite questionable — you could even call them “fake”. So I’d strongly recommend doing thorough data cleaning before training.
Thanks again!
Thanks so much for the quick response — that's incredibly helpful!
Knowing that Unsloth is the foundation gives me a lot of confidence. That alone is a huge insight.
On the dataset side — I'm completely with you. I've noticed the same thing myself. The quality of publicly available distillation datasets is all over the place, and some are honestly questionable at best.
I saw in the README that you used a blend of premium open-source sources from Hugging Face. I'm guessing those are the same ones you used in previous versions? Or did you bring in any new sources for v3?
I'd love to study them myself and do my own data curation — so any pointers on which specific datasets you'd recommend would be hugely appreciated.
Thanks again for taking the time. Big fan of your work — I'll keep following along and supporting!
You sir are the GOAT 🫡
Curious how well this will run as baseline local agent for OpenClaw (simple tasks obviously). My RTX5090 is hungry for models worth running locally :) Hoping for a suitable quant/format soon
Amigo. Felicidades!.
Pregunta. ¿Por qué no juntar Qwopus mas Gemini 3.1 pro?
Qwopus + Gemnini 3.1 Pro en Qwen3.5 4B, 9B y demas? Inclusive en el qwen3.5 0.8B. ¿Por qué?
O no es posible?
El usuario @voidful hizo el finetuning con este modelo: Qwen3.5-9B-gemini-3.1-opus-4.6-reasoning.
Pero creo que con el tuyo es definitivamente mejor con el modelo Qwopus3.5-27B-v3-GGUF.
Piensalo, seria un excelente modelo.
Ademas... Vale la pena que realices el finetuning a los nuevos modelos de Gemma 4 en todas sus versiones.
Thanks so much for the quick response — that's incredibly helpful!
Knowing that Unsloth is the foundation gives me a lot of confidence. That alone is a huge insight.
On the dataset side — I'm completely with you. I've noticed the same thing myself. The quality of publicly available distillation datasets is all over the place, and some are honestly questionable at best.
I saw in the README that you used a blend of premium open-source sources from Hugging Face. I'm guessing those are the same ones you used in previous versions? Or did you bring in any new sources for v3?
I'd love to study them myself and do my own data curation — so any pointers on which specific datasets you'd recommend would be hugely appreciated.
Thanks again for taking the time. Big fan of your work — I'll keep following along and supporting!
I’ve fully open‑sourced my code on GitHub, which you can access through the link on my homepage.
For v3, I introduced new data sources, as publicly available distillation datasets are not very reliable...