Finetuning code and 8b Model?

[mehulsuresh]

Hey we tried this model on some of our datasets and it doesn't really have good one shot performance. Are there plans to release the code to fine-tune in the future? Also will the larger 8b model be released?

[7-Z-7]

Thank you for using it. Can you share examples that didn't work well? I might be able to provide some suggestions. The 8b model can currently be experienced at https://vlac.intern-ai.org.cn/, and API access is planned for the future. Fine-tuning related examples are in plans.

[mehulsuresh]

we were trying to evaluate on out bin picking dataset.
Here are the settings we used
Critic.temperature = 0.5
Critic.top_k = 1
batch_num=5,
ref_num=20,
think=True,
skip=1,
rich=True,
reverse_eval=False,
output_path="results",
fps=float(output_fps),
frame_skip=False,
done_flag=False,
in_context_done=False,
done_threshold=0.95,
video_output=True,

here are some examples.

good2-0-13.mov

bad3-0-26.mov

bad2-0-10.mov

[bbbbbbbbly]

when the finetune code release? the pi 0.6 all ready use the value function to make RL work.

[mehulsuresh]

What pi 0.6 does is more like advantage weighted regression rather than RL. But yeah everybody is working on this now.

…

On Mon, Nov 17, 2025 at 7:39 PM bbbbbbbbly ***@***.***> wrote: *bbbbbbbbly* left a comment (InternRobotics/VLAC#4) <#4 (comment)> when the finetune code release? the pi 0.6 all ready use the value function to make RL work. — Reply to this email directly, view it on GitHub <#4 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEFJGKWRPRPKCESSK2VLY6L35KIG7AVCNFSM6AAAAACHNZOR3GVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTKNBUHA4DIMBYGA> . You are receiving this because you authored the thread.Message ID: ***@***.***>