W
AI-Wiki
← 开源项目

TianyuCodings/NanoJev

已整理
https://github.com/tianyucodings/nanojev

NanoJev 用较小的模型直接为候选动作输出概率,帮助在游戏或决策任务中快速选择下一步,而不需要生成文本答案。它适合想研究并复现实验型决策模型、游戏智能体评测或多候选动作打分流程的开发者和研究者。

项目简介

NanoJev 用较小的模型直接为候选动作输出概率,帮助在游戏或决策任务中快速选择下一步,而不需要生成文本答案。它适合想研究并复现实验型决策模型、游戏智能体评测或多候选动作打分流程的开发者和研究者。

官方仓库资料

主要语言
Python
许可证
MIT
默认分支
main
Stars
2065
Forks
220
Open Issues
7
仓库状态
正常运行
派生仓库
非 Fork
数据采样
2026/9/23 19:06:52
资料抓取
2026/9/23 19:06:52
首次收录
2026/9/23 19:04:11

Topics:仓库未提供

AI 整理 · 基于已抓取的官方资料

核心功能

  • Parallel decision evaluation: batches independent states, questions, and candidate paths in one backbone forward pass.
  • Dynamic candidate choice: returns a probability distribution over 2–255 supplied candidates using a shared scoring head.
  • Boolean prediction: estimates the probability of a proposition with a sigmoid head.
  • Ordered score prediction: returns a distribution and expectation over 2–10 ordered levels.
  • Direct probability outputs: supports ranking, selecting, or sampling actions without generating answer tokens.
  • Unified game checkpoint: the README describes one shared checkpoint used for Maze, Snake, ViZDoom Basic, and ViZDoom Predict Position.
  • Browser replay demos: includes local web replay pages for recorded gameplay comparisons.
  • Public model and dataset: model and data are available from Hugging Face under the unified-games-v1 revision.

适用场景

  • Evaluate candidate actions for game environments such as Maze, Snake, ViZDoom Basic, and ViZDoom Predict Position.
  • Run local browser replays of recorded gameplay comparisons.
  • Experiment with decision-model inference where each request contains a state, a question, and supplied candidates.
  • Download and inspect the public NanoJev checkpoint and associated mixed-task dataset.
  • Reproduce or study the README-described training and evaluation pipeline using the linked documentation.

使用方式

install: - git clone https://github.com/TianyuCodings/NanoJev.git

  • cd NanoJev
  • python -m pip install -r requirements-toy.txt huggingface_hub download_checkpoint_and_data: Use huggingface_hub.snapshot_download to download C-Tianyu/NanoJev and C-Tianyu/NanoJev-Data with revision unified-games-v1 into local checkpoint and data directories. start_inference_service: - python scripts/serve_decisions.py --checkpoint-dir checkpoints/NanoJev-unified --web-root web --port 8765 --disable-native-triton api_endpoint: POST http://127.0.0.1:8765/api/evaluate local_replays: - 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.
  • Open http://127.0.0.1:8080/dev/side-by-side.html?autoplay=1#maze for Maze.

技术特点

  • Repository language is Python.
  • The README describes NanoJev as using Qwen3-0.6B with decision heads.
  • Each request supplies a state, a question, and candidates.
  • Choice uses set attention and a softmax according to the README.
  • Boolean outputs use a sigmoid according to the README.
  • Score returns a probability-weighted level according to the README.
  • unified-games-v1 is the Hugging Face release tag for the matching model and dataset snapshots.
  • hard_lr1e5 is described as the training experiment that produced the selected step-400 checkpoint.
  • The service is started in a CUDA environment according to the quick-start instructions.
  • The README links documentation for release contents, input contract, unified environments, atomic planning, and replay demos.

限制与边界

  • The interactive development site is described as currently requiring access, although recordings can be played locally.
  • The README’s roadmap lists RLCD post-training for broader long-horizon tasks as not yet completed.
  • The README’s roadmap lists shared-prefix inference and larger candidate batches as not yet completed.
  • The README’s roadmap lists broader shooting scenarios and structured input support as not yet completed.
  • Held-out results in the README show NanoJev does not outperform Jev on every task; for example, Maze is listed as 4/10 for NanoJev and 7/10 for Jev.
  • ViZDoom Predict Position success remains limited in the reported test results, with 27/128 successes for NanoJev.

README

The README introduces NanoJev as a 0.6B parallel decision model that outputs complete probability distributions for supplied candidates instead of decoding answer tokens. It reports a unified checkpoint for four game tasks: Maze, Snake, ViZDoom Basic, and ViZDoom Predict Position, with public model and dataset snapshots on Hugging Face under `unified-games-v1`. The project includes quick-start instructions for installing dependencies, downloading checkpoints and data, starting a local inference service, and viewing browser-based recorded game replays. The README also provides reported benchmark results, dataset scale details, release naming notes, development documentation links, and a roadmap for future improvements.

关联 Wiki 词条

暂无已建立的词条关联。