1120_BEHAVIOR_challenge_QwenDual_taskall

StarVLA/1120_BEHAVIOR_challenge_QwenDual_taskall

1120_BEHAVIOR_challenge_QwenDual_taskall is a robotics model from StarVLA. Use it for the robotics task on the model card, and read the license before you ship it in a product.

This repository contains the 100,000-step checkpoint from the 1120BEHAVIORchallengeQwenDualtaskall run. Its public configuration sets datamix: BEHAVIORchallenge and taskid: all. This describes the training scope; it i…

Downloads · 30 days

21% of all-time downloads

All-time downloads

284

Public

Repo size

10 GB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.pt10 GB · 100%

At a glance

Task: Robotics
Access: Public
Created: Nov 27, 2025
Updated: Sep 4, 2026
SHA: ada8847d

Base models

Qwen/Qwen3-VL-4B-Instruct

Model card

From the Hugging Face model README

StarVLA QwenDual for the BEHAVIOR Challenge, all configured tasks (100K)

This repository contains the 100,000-step checkpoint from the 1120_BEHAVIOR_challenge_QwenDual_taskall run. Its public configuration sets data_mix: BEHAVIOR_challenge and task_id: all. This describes the training scope; it is not evidence of an official all-task challenge score.

Model details

Item	Published configuration
Framework	StarVLA `QwenDual`
VLM	Local training snapshot named `Qwen3-VL-4B-Instruct`; revision not recorded
Auxiliary visual encoder	`dinov2_vits14`
Action model	16-layer `DiT-B` flow head: 768 latent width, 12 heads (64 dimensions/head); state/action decoder MLP width 1,024
Action / state dimension	23 / 44
Predicted action chunk	50 steps
Inference diffusion steps	4
Model maximum sequence length	2,048
Image resolution	224 x 224
Dataset mixture / task selector	`BEHAVIOR_challenge` / `all`
Normalization key	`R1Pro`
Uploaded checkpoint	`checkpoints/steps_100000_pytorch_model.pt`

The configuration enables state input but does not specify camera count or ordering. Evaluation must replay the training adapter rather than inventing an observation schema from the 224 x 224 resolution alone.

Training details

Setting	Value in `config.yaml`
Maximum / released step	100,000 / 100,000
Per-device VLA batch size	32
Gradient accumulation	1
Warm-up steps	5,000
Base / interface / action LR	`4e-5` / `1e-5` / `1e-4`
Optimizer	AdamW, betas `(0.9, 0.95)`, epsilon `1e-8`
Scheduler	Cosine with minimum LR `1e-6`
VLA / VLM loss scale	`1.0` / `0.1`
`freeze_modules`	Packaged boolean `true`; the public trainer expects module paths as a string, so this value names/selects no modules
Gradient checkpointing / mixed precision	Enabled / enabled
Seed	42

Files

config.yaml
dataset_statistics.json
summary.jsonl
checkpoints/
└── steps_100000_pytorch_model.pt

The packaged statistics expose one R1Pro normalization entry. Keep the statistics and configuration with the checkpoint when serving it.

Evaluation status

The repository publishes no BEHAVIOR episode manifest or score. In particular, summary.jsonl only lists the 10K checkpoint intervals and is not a benchmark summary. No official-platform, validation-split, or test-split result is assigned to this artifact in this card.

Loading and evaluation

huggingface-cli download StarVLA/1120_BEHAVIOR_challenge_QwenDual_taskall \
  --local-dir 1120_BEHAVIOR_challenge_QwenDual_taskall

export CKPT=$PWD/1120_BEHAVIOR_challenge_QwenDual_taskall/checkpoints/steps_100000_pytorch_model.pt
python deployment/model_server/server_policy.py \
  --ckpt_path "$CKPT" \
  --port 10093 \
  --use_bf16

Use the current StarVLA BEHAVIOR client and validate the server handshake before evaluation. The Hub repository's training helper is not presented here as a reviewed, portable launcher.

Intended use and limitations

This is a research checkpoint for the configured BEHAVIOR challenge mixture. The meaning and coverage of task_id: all depend on the training-time dataset registry; the Hub files do not enumerate the exact task/instance manifest. There is no public result for this checkpoint, no pinned VLM revision, and no evidence of safe transfer to physical robots.

Task: Robotics
Created: Nov 27, 2025
Updated: Sep 4, 2026