sisizhang/swesmith-critic-ppo-tasks

SWE-Smith Critic/PPO 任务包 / Task Packages

中文

本仓库发布 Harbor Agent PPO 项目最终冻结的 1,761 道 Critic 任务目录，以及其中 529 道在线 PPO 训练任务的清单。PPO 任务是 Critic 任务的子集，不重复存放任务包。

Critic train/selection/qualification 分别为 1,424/164/173 题，以 repository 为隔离单位。任务入库前经过结构、补丁、镜像、NOP/Oracle、题面与测试对齐等检查；只有产生完整过程标签轨迹的任务进入这份冻结任务列表。在线 PPO 从 Critic train 集合中另选 529 题。

PPO 训练样本的选择

对每题只统计过程标签完整且终局结果明确的历史 rollout：n 为有效 rollout 数，k 为其中成功数。全选 0 < k < n 的 481 道混合成败题；再从 k = 0 的始终失败题和 k = n 的始终成功题各选 24 道，共 529 题。多数混合成败题来自四次历史 rollout 中通过 1、2 或 3 次的任务。

两端抽样使用冻结种子 31001，按 sha256(seed|split|side|task_id) 排序，固定选题结果。data/ppo_train.jsonl 逐题给出原始冻结 ID、对应的任务目录、n、k 与组别；529 道任务的完整文件位于本仓库 tasks/train/ 下，不另存重复副本。

文件结构

data/
  critic_train.jsonl
  critic_selection.jsonl
  critic_qualification.jsonl
  ppo_train.jsonl
tasks/
  train/<task_id>/
  selection/<task_id>/
  qualification/<task_id>/
README.md

每条 JSONL 至少提供 task_id、split、task_path 和来源批次；PPO 清单还提供 canonical_task_id、n、k 和 selection_group。task_path 是相对于本仓库根目录的路径。每个任务目录包含 task.toml 及其任务说明、环境、测试和 Oracle 解法文件。只用 datasets 加载 JSONL 不会自动下载完整任务包；执行任务请用 snapshot_download 获取整个仓库。

本仓库只发布任务目录与任务选择元数据，不包含 Critic 的 6,119 条原始 process-complete 轨迹、逐 token 奖励标签、Harbor 原始作业、任务镜像或私有运行凭据。任务数不能当作轨迹数或模型更新样本数。Critic 训练时另有 65,536 token 上下文筛选，实际有效轨迹数与冻结来源数不同。

使用与边界

任务包包含参考补丁、隐藏测试相关材料和 Oracle 信息。它们公开后不能再作为未公开盲测集；训练与评测应按 split 隔离，并避免把答案或 verifier 证据注入 Agent 提示词。容器镜像未随仓库发布，完整运行还需要兼容的 Harbor 沙箱和 verifier。

任务材料来自不同上游仓库，统一标注为 license: other；使用者须分别核对源任务和代码仓库的许可证。模型来源与评测口径见 项目 Space。

支持该项目

数据、开发、GPU和CPU等训练的费用均由我自行承担。如果您觉得这个模型对您有用，并希望资助更大的数据集和未来的开源版本，可以给我点赞，留言。

English

This repository releases 1,761 frozen Critic task directories and a list of 529 online PPO training tasks. The PPO tasks are a subset of the Critic tasks; their packages are not duplicated.

The Critic train/selection/qualification splits contain 1,424/164/173 tasks, separated by upstream repository. Tasks pass structure, patch, image, NOP/Oracle, and problem/test-alignment checks before inclusion; only tasks yielding complete process-labeled trajectories enter this frozen list. The 529 PPO tasks are selected from Critic train.

PPO task selection

Only historical rollouts with complete process labels and a definite terminal outcome are counted: n is the number of valid rollouts and k the number that succeeded. We include all 481 mixed-outcome tasks with 0 < k < n, plus 24 always-failing (k = 0) and 24 always-successful (k = n) tasks, for 529 total. Most mixed-outcome tasks succeeded in one, two, or three of four historical rollouts.

Both endpoint groups use frozen seed 31001 and are ordered by sha256(seed|split|side|task_id) for reproducible selection. data/ppo_train.jsonl lists each frozen ID, task directory, n, k, and selection group. The full task files are under tasks/train/.

Files

data/
  critic_train.jsonl
  critic_selection.jsonl
  critic_qualification.jsonl
  ppo_train.jsonl
tasks/
  train/<task_id>/
  selection/<task_id>/
  qualification/<task_id>/
README.md

Each JSONL row includes at least task_id, split, task_path, and source batch; the PPO list also includes canonical_task_id, n, k, and selection_group. task_path is relative to the repository root. Each task directory contains task.toml, instructions, environment files, tests, and an Oracle solution. Loading only the JSONL through datasets does not fetch the task packages; use snapshot_download for the full repository.

This repository contains task directories and selection metadata. The 6,119 process-complete Critic trajectories are released separately; original Harbor jobs, task images, and private credentials are not included. Task counts are not trajectory or model-update counts. Critic training applies a separate 65,536-token context filter.

Use and limitations

Task packages contain reference patches, hidden-test-related material, and Oracle information. Once public, they are not an unseen blind test set. Keep splits separate and do not expose answers or verifier evidence to the evaluated Agent. Container images are not included; full execution requires a compatible Harbor sandbox and verifier.

The materials come from multiple upstream repositories and are marked license: other; users should check the relevant source-task and code licenses. See the project Space for model provenance and results.

Support This Project

I personally cover the costs of data, development, GPU and CPU training. If you find this model useful and would like to support larger datasets and future open-source releases, please give it a like and leave a comment.