TianyuCodings/NanoJev

选用指南

AI 根据 README 整理 · 2026/10/17 05:15:54

使用方式

通过命令行启动推理服务，发送状态和问题批次到 HTTP API，或本地运行网页演示。

查看 README 依据

repo_id="C-Tianyu/NanoJev", revision="unified-games-v1", local_dir="checkpoints/NanoJev-unified", allow_patterns=["best.safetensors", "config.json", "tokenizer/*", "backbone_config/*"], ) snapshot_download( repo_id="C-Tianyu/NanoJev-Data", repo_type="dataset", revision="unified-games-v1", local_dir="data/NanoJev-unified", ) ``` Start inference in a CUDA environment: ```bash python scripts/serve_decisions.py \ --checkpoint-dir checkpoints/NanoJev-unified \ --web-root web --port 8765 --disable-native-triton ```

The service loads the model once. Send state/question batches to **`POST http://127.0.0.1:8765/api/evaluate`**. To explore the recorded games locally: ```bash python3 -m http.server 8080 --bind 127.0.0.1 --directory web ``` Open **http://127.0.0.1:8080/dev/?autoplay=1** for ViZDoom Basic or **http://127.0.0.1:8080/dev/side-by-side.html?autoplay=1#maze** for Maze. ## Development notes

部署要求

需要 CUDA 环境以运行推理服务。

查看 README 依据

repo_id="C-Tianyu/NanoJev", revision="unified-games-v1", local_dir="checkpoints/NanoJev-unified", allow_patterns=["best.safetensors", "config.json", "tokenizer/*", "backbone_config/*"], ) snapshot_download( repo_id="C-Tianyu/NanoJev-Data", repo_type="dataset", revision="unified-games-v1", local_dir="data/NanoJev-unified", ) ``` Start inference in a CUDA environment: ```bash python scripts/serve_decisions.py \ --checkpoint-dir checkpoints/NanoJev-unified \ --web-root web --port 8765 --disable-native-triton ```

已知限制

当前模型在 ViZDoom Predict Position 任务上测试成功率为 27/128，低于 Basic 任务的 128/128。

查看 README 依据

Successful episodes on the complete **274-case test set**, using the same observation interface, candidate actions and seeded epsilon-greedy controller across systems: | Model | Maze | Snake | Basic | Predict Position | |---|---:|---:|---:|---:| | **NanoJev** | **4/10** | **8/8** | **128/128** | **27/128** | | [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) | 7/10 | 8/8 | 56/128 | 11/128 | | Untuned Qwen3-0.6B | 2/10 | 0/8 | 56/128 | 11/128 |