RynnValue-8B

RynnValue-8B uses RynnBrain-8B and fixed bins to score a video trajectory against a task instruction. The heads estimate time remaining and signed time displacement. DAMO reports a mean Kendall τₐ of 0.675 on RBM-EVAL-OOD policy ranking with a causal mask.

Vollständige Modellbeschreibung ansehen

Publikationen

Publikationen konnten nicht geladen werden. Erneut versuchen

Zugang und Nutzung

Öffentliche Code-Ressourcen
Öffentlicher Zugang
Modelldateien
Öffentlicher Zugang

Der Trainingscode dieser Version ist unbestätigt. Öffentliche Code-Ressourcen belegen nicht die Verfügbarkeit der genauen Trainingsimplementierung. Prüfhinweise ansehen

Trainings- / Fine-Tuning-Code

Modelldateien

Modelldateien
RynnValue-8B (native)ZugangGewichtelizenzKommerzielle Nutzung
Unter BedingungenBedingungen

Geprüft am · 2026-11-19

Bedingungen

Apache-2.0 applies. When redistributing, include the license and retain applicable notices.

RynnValue-8B (LeRobot)ZugangGewichtelizenzKommerzielle Nutzung
Unter BedingungenBedingungen

Geprüft am · 2026-11-19

Bedingungen

Apache-2.0 applies. When redistributing, include the license and retain applicable notices.

Veröffentlichungsangaben

QuelleGeprüft am · 2026-11-19

Vor der Bereitstellung prüfen

Verfügbare Gewichte oder Code belegen keine sofortige Roboterkompatibilität. Prüfen Sie Konfiguration, Daten und Schnittstellen in der Quelldokumentation.

Unterlagen zu Bereitstellung und Nutzung

Laufzeit und Frameworks
LeRobot / PyTorch / Transformers

Architektur und Ein-/Ausgaben

Modelltyp
Belohnung und Bewertung
Architektur
Transformer
Backbone
RynnBrain-8B (Qwen3-VL)
Gewichteparameter
9.575 B
Eingabe
Sprache / Video

Einsatzbereich und Training

Implementierungsquelle

Offiziell

Funktionsbereich

Reward and progress

Ausgabe

Belohnung / Fortschritt

Pipeline-Aufgabe

Robotic reward model

Vorhersageausgabe

Time to completion, temporal displacement and task verification

Aufgabenumfang

Mehrere Aufgaben

Ressourcen

Alle Parameter und Prüfhinweise (Modellkarte)

Fixed-bin 8B and its two weight packages

The official 8B name differs from the stored tensor count: 9,574,961,904 BF16 elements in each package. The upstream model uses RynnBrain-8B on Qwen3-VL and fixed 256-bin temporal value tokenization. The official model zoo reports RBM-EVAL-OOD Kendall τₐ of 0.675 for this fixed-bin model. The native weight upload was July 27, 2026 at 21:59:15 Taipei time. LeRobot created its repository on November 19 at 23:14:53 and uploaded the conversion at 23:15:23 Taipei time. Both are public downloads under Apache-2.0. The original conversion was based on upstream revision 12848f45ef339e95671c89bc4a3e7b21b096d136. Its single safetensors file is 19,150,038,712 bytes with 890 stored tensors.

LeRobot inference

Use LeRobot source with RynnValue integration PR #4555. For a pinned load, call RewardModelConfig.from_pretrained("lerobot/RynnValue-8B", revision="83c91ad43cf29deccd432d8ddd37833f638c878a"), then pass the config to make_reward_model and the preprocessor factory. The config, weights and processor/ load from that commit. The updated config points to the LeRobot repository. The config at the older commit 0a24577 has model_id and model_revision fields that this integration rejects. This package supports inference only. Use predict_remaining_time(batch).remaining_time_s to estimate seconds remaining or predict_success(batch).success for the model’s task-success judgment. Offline scoring defaults to inference_fps=1.0, or about one query per second, and includes the first and last trajectory frame. It samples eight evenly spaced frames from the prefix available at each scoring point, repeating frames for short prefixes. It linearly interpolates the scores between points. Image long edges are capped at 640 pixels. The default image and instruction keys are observation.images.top and task. Default metadata prompting needs a robot or camera description. The integration source requires Python ≥3.12 and PyTorch ≥2.7 and <2.16. The rynnvalue extra requires Transformers ≥5.22.0 and <5.23.0. Sources: Native upload, LeRobot conversion, implementation, model zoo. Cover: Figure 1 by Huang et al., https://arxiv.org/abs/2608.09853v2, CC BY 4.0, resized and padded. Configuration update Current config file Current integration dependencies

Native inference and official mirror

DAMO’s native toolkit requires Python ≥3.10, PyTorch ≥2.4.0 and Transformers ≥4.57.0 and <5. The official example loads the custom model and AutoProcessor with trust_remote_code=True, then calls processor.process_episode on the task instruction and sampled images. DAMO suggests a single GPU with at least 24 GB of memory for 8B BF16. The four native safetensors shards have matching sizes and published SHA-256 hashes on ModelScope and Hugging Face. See the native dependency manifest. The native forward example averages out.value.pred_value across value heads for per-frame remaining seconds. The native demo also parses generated task-match and success answers.

Quelle oder Korrektur einreichen