liangzhidanta/claude-code-glm53-swesmith-compact-trajectories

Claude Code GLM-5.3 SWE-smith Compact-Aware Trajectories

Targeted SFT data for native context compaction and post-compaction continuation in coding agents.

English | 简体中文

---

English

Dataset at a Glance

189 execution-verified compact-aware source trajectories → context-correct conversion → 890 context-correct training segments, of which the recommended Compact-Focused training units comprise 626 Compact Summary segments and 17 long-context auxiliary segments (the remaining 267 are optional pre-compact segments).

189 source trajectories ≠ 890 context-correct segments ≠ 643 compact-focused training segments. These counts refer to different units of analysis. The 189 source sessions are represented as 890 training segments because native compaction introduces model-visible context boundaries — this is not data augmentation.

189 source trajectories
├── 138 trajectories with 1 native compaction
└──  51 trajectories with 2 native compactions
     → 138×1 + 51×2 = 240 native compact events

Each compact event produces one Compact Summary supervision segment (real CC forwarded request → GLM-5.3 summary). Each post-compact continuation is reconstructed from the actual model-visible wire request as Post-Compaction Continuation segments. Because Claude Code may perform mid-stretch trimming, one source trajectory can yield multiple continuation segments.

189 source trajectories
→ 240 Compact Summary segments + 403 Post-Compaction Continuation segments
= 643 Compact-Focused training segments
→ 626 deployment-aligned (≤40,960 tokens) + 17 long-context auxiliary

These are not 890 independent SWE tasks — they are context-correct training views derived from 189 successful sessions on the same task pool as V1.

Two Types of Compact-Aware Supervision

1. Compact Summary Segments

Input: the real Claude Code native compact FORWARDED request (tools stripped, thinking disabled, flattened history, summary guard). Target: the GLM-5.3 faithful state summary.

These segments train the model to:

• Compress the current agent working state
• Distinguish completed vs. not-completed actions
• Preserve actual file modification state
• Preserve test execution state and results
• Retain current hypothesis, pending work, and next action
• Avoid fabricated completion claims
• Avoid verbatim transcription of tool output

2. Post-Compaction Continuation Segments

Input: the real model-visible context of Claude Code's next request after compaction (context reset — summary + retained turns, no pre-compaction history). Target: subsequent assistant reasoning, text, and tool calls.

These segments train the model to continue completing a coding task using only the compressed state.

The two types address different problems: compression quality and continuation after compression.

Why Split the 189 Full Trajectories?

Real Claude Code native compaction follows:

pre-compaction context → summary → context reset → continuation

During continuation, the model cannot see the full pre-compaction history that was trimmed. If we naively concatenate a full trajectory end-to-end into one training sequence, post-compaction targets would attend to pre-compaction context that is invisible at inference time — a train–inference mismatch.

Our audit of the legacy full-concat approach found:

• Post-c1 targets leaked pre-compact tokens: mean ≈ 9,895 (66% of their training prefix)
• Post-c2 targets leaked: mean ≈ 28,895 (81% of their prefix)

We therefore do not split to "expand" 189 into more data. We split to reproduce the actual model-visible context at native compaction boundaries.

Core Statistics

Metric	Value
Source trajectories	189
1-compaction trajectories	138
2-compaction trajectories	51
Native compact events	240
Compact-summary segments	240
Post-compaction continuation segments	403
Optional pre-compaction segments	247
Context-correct segments total	890
Compact-focused segments	643
Deployment-aligned ≤40k segments	626
Long-context auxiliary segments	17
Unique repos	77
Train / validation tasks	173 / 16
Median segment tokens	23,135
P95 segment tokens	34,355
Teacher	GLM-5.3
Harness	Claude Code 2.1.258
Collection context	40,960 CC-visible

Data Construction

Data	Purpose	Recommended for SFT	Preserves full session	Context-correct
Full Source Trajectories (189)	Provenance / audit / research	NO	YES	N/A
Context-Correct Segments (890)	SFT training	YES	NO (split at boundaries)	YES
Compact-Focused 40k (626 V2)	Current recommended recipe	YES	NO	YES (≤40,960 tokens)
Long-Context Aux (17)	Valid but >40k	Future use	NO	YES (>40,960 tokens)

⚠️ DO NOT naively concatenate full source trajectories for SFT

Native compaction resets the model-visible context. Full-session concatenation causes train–inference mismatch (see audit numbers above).

Dataset Structure

full_source/                    # 189 original sessions (provenance, audit)
context_correct/
  compact_summary_*.jsonl       # 240 segments (compact request → summary)
  post_compact_*.jsonl          # 403 segments (post-compact visible context → continuation)
  optional_precompact_*.jsonl   # 247 segments (pre-compact agent behavior)
compact_focused_40k/            # 626 V2 + V1 replay = recommended training view
long_context_aux/               # 17 segments >40,960 tokens
manifests/                      # source, segment, and quality manifests
metadata/                       # stats, provenance, split manifest
assets/                         # statistics dashboard
tools/                          # reproducible figure generation script

Schema (slime-ready)

{
  "messages": [...],
  "tools": [...],
  "metadata": {
    "sample_id": "task_id::segment_kind",
    "task_id": "...",
    "segment_type": "compact_summary | post_compact | pre_compact",
    "source_trajectory_id": "...",
    "compact_count": 1
  }
}

Mask semantics (Qwen3, slime qwen3 loss-mask):

• System / user / tool_result → mask 0
• Assistant reasoning / text / tool_call → mask 1
• Compact request context → mask 0
• GLM compact summary response → mask 1
• Post-compact assistant → mask 1

Recommended Usage

This repo provides compact-aware targeted data only. Our SFT-v2 training recipe:

V1 normal-agent dataset (1003 trajectories)
+ V2 compact-summary segments (240)
+ V2 post-compaction continuation segments (403)

We do not copy V1 data into this repo. Users should combine it with V1 as needed.

• compact_focused_40k/ — recommended training view (deployment-aligned)
• full_source/ — provenance and research only (do not train on naive concat)
• optional_precompact_*/ — for ablation
• long_context_aux/ — for >40k context research

Relationship to V1

	V1	V2 (this repo)
Focus	Normal coding agent policy	Compaction + post-compaction continuation
Tasks	Same SWE-smith pool	Same pool, re-collected
Trajectories	1003 (no compaction)	189 (native compaction)
Value	Base coding capability	Targeted compaction capability

V2 is a targeted capability data extension, not a task expansion of V1.

Limitations

• Task pool is shared with V1 (same-task, different-trajectory supervision)
• Compact summaries generated by GLM-5.3 may contain residual errors
• Collection protocol: single 40k window; behavior at other context sizes may differ
• 17 segments exceed 40,960 tokens and are provided as auxiliary

Citation

If you use this dataset, please cite both V1 and V2:

@dataset{v1_claude_code_glm53_swesmith,
  title={Claude Code GLM-5.3 SWE-smith Trajectories},
  author={liangzhidanta},
  year={2026},
  url={https://huggingface.co/datasets/liangzhidanta/claude-code-glm53-swesmith-trajectories}
}
@dataset{v2_compact_aware,
  title={Claude Code GLM-5.3 SWE-smith Compact-Aware Trajectories},
  author={liangzhidanta},
  year={2026},
  url={https://huggingface.co/datasets/liangzhidanta/claude-code-glm53-swesmith-compact-trajectories}
}

---

简体中文

数据集概览

189 条执行验证的 compact-aware 源轨迹 → 上下文正确转换 → 890 个上下文正确训练片段，其中推荐的 Compact-Focused 训练单元包含 626 个 Compact Summary 片段 和 17 个长上下文辅助片段（其余 267 个为可选的 pre-compact 片段）。

189 条源轨迹 ≠ 890 个上下文正确训练片段 ≠ 643 个 compact-focused 训练片段。 三者对应不同统计单位。189 条源会话依据真实 native compaction 产生的模型可见上下文边界，被表示为 890 个训练片段——这不是人工数据增强。

189 条源轨迹
├── 138 条经历 1 次 native compaction
└──  51 条经历 2 次 native compaction
     → 138×1 + 51×2 = 240 次 native compact 事件

每次 compact 事件产生一个 Compact Summary 监督片段（真实 CC 转发请求 → GLM-5.3 摘要）。每次 compact 后的 Agent 续跑从真实模型可见 wire request 重建为 Post-Compaction Continuation 片段。由于 Claude Code 可能进行中段裁剪，一条源轨迹可能对应多个续跑片段。

189 条源轨迹
→ 240 个 Compact Summary 片段 + 403 个 Post-Compaction Continuation 片段
= 643 个 Compact-Focused 训练片段
→ 626 个 deployment-aligned（≤40,960 tokens）+ 17 个长上下文辅助

这些不是 890 个独立 SWE 任务——它们是从 189 条成功会话（与 V1 使用同一任务池）导出的上下文正确训练视图。

两类压缩感知监督数据

1. Compact Summary 片段

输入： 真实 Claude Code native compact 转发请求（工具已剥离、thinking 已禁用、历史已扁平化、含 summary guard）。 目标： GLM-5.3 忠实状态摘要。

这些片段训练模型：

• 压缩当前 Agent 工作状态
• 区分已完成 / 未完成操作
• 保留真实文件修改状态
• 保留测试执行状态与结果
• 保留当前假设、待办任务、下一步行动
• 避免编造完成状态
• 避免逐字转录工具输出

2. Post-Compaction Continuation 片段

输入： compact 后 Claude Code 下一次模型调用的真实可见上下文（上下文重置——摘要 + 保留轮次，无 pre-compact 历史）。 目标： 后续 assistant 推理、文本和工具调用。

这些片段训练模型"仅使用压缩后的状态继续完成 Coding Task"。

两类数据解决不同问题：压缩质量与压缩后的续跑能力。

为什么要拆分这 189 条完整轨迹？

真实 Claude Code native compaction 遵循：

压缩前上下文 → 摘要 → 上下文重置 → 续跑

续跑期间，模型无法看到已被裁剪的完整压缩前历史。如果将完整轨迹首尾拼接为一个训练序列，压缩后目标将注意到推理时不可见的压缩前上下文——形成训练—推理上下文不一致。

我们对 legacy 全拼接方案的审计发现：

• post-c1 目标平均泄漏压缩前 tokens：约 9,895（训练前缀的 66%）
• post-c2 目标平均泄漏：约 28,895（前缀的 81%）

因此我们不是为了"把 189 条变成更多数据"而人为拆分，而是为了在 native compaction 边界处重现真实的模型可见上下文。

核心统计

指标	数值
源轨迹	189
1 次压缩轨迹	138
2 次压缩轨迹	51
Native compact 事件	240
Compact-summary 片段	240
Post-compact 续跑片段	403
可选 pre-compact 片段	247
上下文正确片段总数	890
Compact-focused 片段	643
Deployment-aligned ≤40k 片段	626
长上下文辅助片段	17
唯一仓库	77
训练 / 验证任务	173 / 16
片段 token 中位数	23,135
P95 片段 tokens	34,355
教师	GLM-5.3
运行时	Claude Code 2.1.258
采集上下文	40,960 CC-visible

数据构造

数据	用途	推荐用于 SFT	保留完整会话	上下文正确
Full Source Trajectories（189）	溯源 / 审计 / 研究	否	是	不适用
Context-Correct Segments（890）	SFT 训练	是	否（边界拆分）	是
Compact-Focused 40k（626 V2）	当前推荐配方	是	否	是（≤40,960 tokens）
Long-Context Aux（17）	有效但 >40k	未来使用	否	是（>40,960 tokens）

⚠️ 不要将 full source trajectories 直接拼接用于 SFT

Native compaction 会重置模型可见上下文。全拼接导致训练—推理不一致（见上方审计数字）。

数据结构

full_source/                    # 189 条原始会话（溯源、审计）
context_correct/
  compact_summary_*.jsonl       # 240 片段（compact 请求 → 摘要）
  post_compact_*.jsonl          # 403 片段（compact 后可见上下文 → 续跑）
  optional_precompact_*.jsonl   # 247 片段（compact 前 Agent 行为）
compact_focused_40k/            # 626 V2 + V1 replay = 推荐训练视图
long_context_aux/               # 17 片段 >40,960 tokens
manifests/                      # 源、片段、质量 manifest
metadata/                       # 统计、溯源、split manifest
assets/                         # 统计图
tools/                          # 可复现图表生成脚本

Schema（slime-ready）

{
  "messages": [...],
  "tools": [...],
  "metadata": {
    "sample_id": "task_id::segment_kind",
    "task_id": "...",
    "segment_type": "compact_summary | post_compact | pre_compact",
    "source_trajectory_id": "...",
    "compact_count": 1
  }
}

Mask 语义（Qwen3，slime qwen3 loss-mask）：

• System / user / tool_result → mask 0
• Assistant reasoning / text / tool_call → mask 1
• Compact 请求上下文 → mask 0
• GLM compact 摘要响应 → mask 1
• Post-compact assistant → mask 1

推荐使用方法

本仓库仅提供 compact-aware 针对性数据。我们的 SFT-v2 训练配方：

V1 normal-agent 数据集（1003 条轨迹）
+ V2 compact-summary 片段（240 个）
+ V2 post-compaction 续跑片段（403 个）

我们不将 V1 数据复制入本仓库。用户应按需与 V1 组合使用。

• compact_focused_40k/ — 推荐训练视图（deployment-aligned）
• full_source/ — 仅用于溯源和研究（不要直接拼接训练）
• optional_precompact_*/ — 用于消融实验
• long_context_aux/ — 用于 >40k 上下文研究

与 V1 的关系

	V1	V2（本仓库）
定位	基础 Coding Agent 策略	压缩 + 压缩后续跑
任务	同一 SWE-smith 池	同一池，重新采集
轨迹	1003 条（无压缩）	189 条（native compaction）
价值	基础编码能力	针对性压缩能力

V2 是针对性能力数据扩展，不是 V1 的任务扩容。

局限性

• 任务池与 V1 共享（同任务、不同轨迹的监督）
• GLM-5.3 生成的 compact 摘要可能存在残余错误
• 采集协议：单一 40k 窗口；其他上下文大小下的行为可能不同
• 17 个片段超过 40,960 tokens，作为辅助提供

引用

如使用本数据集，请同时引用 V1 和 V2：

@dataset{v1_claude_code_glm53_swesmith,
  title={Claude Code GLM-5.3 SWE-smith Trajectories},
  author={liangzhidanta},
  year={2026},
  url={https://huggingface.co/datasets/liangzhidanta/claude-code-glm53-swesmith-trajectories}
}
@dataset{v2_compact_aware,
  title={Claude Code GLM-5.3 SWE-smith Compact-Aware Trajectories},
  author={liangzhidanta},
  year={2026},
  url={https://huggingface.co/datasets/liangzhidanta/claude-code-glm53-swesmith-compact-trajectories}
}