google/medgemma-1.5-4b-it
Image-Text-to-Text
โข
4B
โข
Updated
โข
91.4k
โข
351
Alfaxad Eyembe PRO
Alfaxad
AI & ML interests
AI, Robotics
Recent Activity
liked
a model
about 23 hours ago
google/medgemma-1.5-4b-it
liked
a dataset
2 days ago
nsk7153/MedCalc-Bench-Verified
liked
a dataset
11 days ago
Atnafu/Afri-MCQA
Organizations
liked
a
model
about 23 hours ago
liked
a
dataset
2 days ago
liked
a
dataset
11 days ago
liked
a
dataset
13 days ago
upvoted
2
collections
17 days ago
liked
a
dataset
20 days ago
liked
a
dataset
24 days ago
liked
a
Space
about 1 month ago
liked
5
datasets
about 1 month ago
reacted to
danielhanchen's
post with ๐ค
about 1 month ago
Post
5495
NVIDIA releases Nemotron 3 Nano, a new 30B hybrid reasoning model! ๐ฅ
Has 1M context window & best in class performance for SWE-Bench, reasoning & chat. Run the MoE model locally with 24GB RAM.
GGUF: unsloth/Nemotron-3-Nano-30B-A3B-GGUF
๐ Step-by-step Guide: https://docs.unsloth.ai/models/nemotron-3
Has 1M context window & best in class performance for SWE-Bench, reasoning & chat. Run the MoE model locally with 24GB RAM.
GGUF: unsloth/Nemotron-3-Nano-30B-A3B-GGUF
๐ Step-by-step Guide: https://docs.unsloth.ai/models/nemotron-3
- 1 reply
liked
a
dataset
about 1 month ago
liked
2
models
about 1 month ago
updated
a
model
about 1 month ago
reacted to
IliaLarchenko's
post with ๐ฅ
about 1 month ago
Post
1142
๐ BEHAVIOR Challenge 1st Place โ Solution Summary
My team recently won 1st place in the BEHAVIOR Challenge at NeurIPS.
The competition focused on training a single policy to complete 50 long-horizon household tasks in simulation.
We built an end-to-end policy based on Pi0.5 with a bunch of custom modifications. Everything is open-sourced, and it should be useful for anyone exploring VLAs or adapting them to specific tasks.
Key Architecture Changes:
- Replaced language model with 50 trainable task embeddings (no text at all)
- Correlated noise for Flow Matching: ฯต โผ N(0, 0.5I + 0.5ฮฃ) using dataset action covariance
- Learnable mixed-layer attention: each action expert layer attends to a trainable mix of all VLM layers
- System 2 stage tracking: model predicts task stage, we smooth it with voting and feed it back as context
Training:
- Multi-sample Flow Matching: 15 FM samples per VLM pass to reduce gradient variance
- Delta action space + per-timestamp normalization
- FAST auxiliary loss and stage prediction loss
- Trained on 224ร224 RGB + proprioception only
- We use 4 fine-tuned checkpoints, all derived from a multi-task model trained on all 50 tasks
Inference Optimizations:
- Soft inpainting: predict 30 actions, execute 26, use 4 as an input for the next chunk
- Correlation-aware guidance of inpainting to keep action chunks smooth
- 1.3ร speedup via cubic spline compression
- General correction rule: reopen gripper after failed grasps
๐ Code and Models:
- Code: https://github.com/IliaLarchenko/behavior-1k-solution
- Weights: IliaLarchenko/behavior_submission
- Paper: Task adaptation of Vision-Language-Action model: 1st Place Solution for the 2025 BEHAVIOR Challenge (2512.06951)
My team recently won 1st place in the BEHAVIOR Challenge at NeurIPS.
The competition focused on training a single policy to complete 50 long-horizon household tasks in simulation.
We built an end-to-end policy based on Pi0.5 with a bunch of custom modifications. Everything is open-sourced, and it should be useful for anyone exploring VLAs or adapting them to specific tasks.
Key Architecture Changes:
- Replaced language model with 50 trainable task embeddings (no text at all)
- Correlated noise for Flow Matching: ฯต โผ N(0, 0.5I + 0.5ฮฃ) using dataset action covariance
- Learnable mixed-layer attention: each action expert layer attends to a trainable mix of all VLM layers
- System 2 stage tracking: model predicts task stage, we smooth it with voting and feed it back as context
Training:
- Multi-sample Flow Matching: 15 FM samples per VLM pass to reduce gradient variance
- Delta action space + per-timestamp normalization
- FAST auxiliary loss and stage prediction loss
- Trained on 224ร224 RGB + proprioception only
- We use 4 fine-tuned checkpoints, all derived from a multi-task model trained on all 50 tasks
Inference Optimizations:
- Soft inpainting: predict 30 actions, execute 26, use 4 as an input for the next chunk
- Correlation-aware guidance of inpainting to keep action chunks smooth
- 1.3ร speedup via cubic spline compression
- General correction rule: reopen gripper after failed grasps
๐ Code and Models:
- Code: https://github.com/IliaLarchenko/behavior-1k-solution
- Weights: IliaLarchenko/behavior_submission
- Paper: Task adaptation of Vision-Language-Action model: 1st Place Solution for the 2025 BEHAVIOR Challenge (2512.06951)