Valen-Preview-0923
✨ Introduction
Valen brings visual perception to System One decision-making. Inspired by Jev, it evaluates text, images and video against task instructions and returns probabilities over supplied candidates, without generating answer tokens.
This repository hosts the 2B Sokoban-trained preview, with the same checkpoint as Valen-Sokoban-RLCD-2B. It builds on Qwen3.5-2B and supports the project's Choice, Noul and Score decision interface. The Valen name comes from Leigh Van Valen's Red Queen hypothesis: continual evolution is necessary to keep pace.
Training starts from the General 100k SFT decision head, then runs RLCD on 30,000 Sokoban records for three epochs using vision_top (90,000 record exposures). The checkpoint includes the decision head, language-model LoRA, visual merger and final four visual blocks, plus optimizer and training state. It does not duplicate the frozen backbone; both downloads below are required.
🎬 Demos
Against a 27B generation model
On the same Sokoban level with a 10-step cap, Valen-Preview-0923 solves the puzzle in 9 decisions with 1.13 seconds of cumulative decision latency. Qwen3.8-27B-FP8 takes 198.05 seconds in thinking mode and fails to solve it in no-thinking mode. The thinking replay runs at 20× speed; the other two run at 1×.
Image blur and action confidence
Nine Gaussian blur levels show how Valen-Preview-0923 changes its action probabilities and decision confidence as visual information becomes less clear. On the clear image, the optimal action DOWN has 93.7% probability and decision confidence is 91.6%; at 72 px blur, these values are 39.4% and 19.2%.
Four games in parallel
Four successful Valen-Preview-0923 trajectories run side by side at 1× speed with no playback acceleration. Each game takes 7–10 decisions, averaging 122–128 ms per step, and all four finish within 1.24 seconds.
📥 Model downloads
| Model | Role | Download |
|---|---|---|
| Valen-Preview-0923 | Trained Valen checkpoint | 🤗 This repository |
| Qwen3.5-2B | Frozen base model | 🤗 Hugging Face |
Required base revision: 15852e8c16360a2fea060d615a32b45270f8a8fc. The setup script in Quick start downloads and verifies that revision.
Datasets: General 100k training · General 5k evaluation · Sokoban training and evaluation.
📊 Evaluation results
The following results apply to this Preview checkpoint:
| Evaluation | Result |
|---|---|
| 500 single-step Sokoban action questions | 87.60% accuracy |
| 100 complete Sokoban games, at most 200 moves | 38/100 solved |
These are the selected simple Sokoban levels: 35 easy and 65 medium. The set retained all successful 2B RLCD cases before filling to 100, so it is model-outcome-conditioned, not an unbiased full benchmark. Single-step accuracy and complete-game success are different metrics.
The figure below provides context across the Valen model family. General is on the left and Sokoban is on the right; each panel uses separately trained checkpoints. The General results are not measurements of this Sokoban preview. RL denotes RLCD, and the Preview corresponds to Valen-2B(RL) in the Sokoban panel.
Full training details, Model Cards, loss curves and timing conditions are in the technical notes.
🚀 Quick start
Use Linux, Python 3.10+ and an NVIDIA GPU. Install the Valen environment, then download both the Preview checkpoint and the matching base model:
git clone https://github.com/Liuziyu77/Valen.git Valen
cd Valen
bash scripts/setup/bootstrap.sh
source .venv/bin/activate
hf download Valen-Team/Valen-Preview-0923 --local-dir models/Valen-Preview-0923
python scripts/setup/prepare_model.py
python -m valen.inference \
--checkpoint models/Valen-Preview-0923 \
--data data/smoke/train.jsonl \
--output predictions.jsonl
data/smoke contains synthetic examples for checking the pipeline. Replace it with your own JSONL file using the documented record format for your task.
The published config.json uses model_path=models/Qwen3.5-2B and stage=vision_top. Run from the Valen repository root and keep that stage so all trained visual and LoRA parameters are loaded. If you store the base model elsewhere, change only model_path in your local copy. This is the project's checkpoint.pt + config.json format, not an AutoModel.from_pretrained repository.
Checkpoint provenance
The published checkpoint.pt and config.json are copied byte-for-byte from the completed three-epoch run. Training details and parameter counts are in training_metadata.json.
Checkpoint SHA-256:
836622efe78fe757e2627aa6050c223d461c424ea430a1c110c15e6f42f5a012
🤝 Contributions
Contributions to Valen are welcome. Open an issue to report a problem, share a use case or discuss experimental results. Submit a pull request to improve the code or documentation, contribute training data or add evaluation tasks.
Scan the QR code below to join the Valen WeChat group, discuss the project and share your experiments.
📄 License and acknowledgments
This repository retains its Apache 2.0 license. The Qwen3.5 base model and source datasets follow their respective licenses. Built on Qwen3.5, with the decision interface inspired by TypeSafe's Jev.