| --- |
| license: mit |
| language: |
| - en |
| pretty_name: DistDebug-Bench Traces |
| size_categories: |
| - 1K<n<10K |
| tags: |
| - distributed-systems |
| - debugging |
| - root-cause-analysis |
| - agents |
| - trajectories |
| - software-engineering |
| configs: |
| - config_name: attempts |
| data_files: |
| - split: train |
| path: data/attempts.parquet |
| default: true |
| --- |
| |
| # DistDebug-Bench Traces |
|
|
| Agent traces behind the main results (Table 1) of *DistDebug-Bench: Can LLM |
| Agents Diagnose Bugs in Distributed Systems?* (NeurIPS 2026, Evaluations and |
| Datasets Track): 17 agent configurations, 120 bugs each, and the three |
| attempts per bug that the paper scores, for 6,120 attempts in total. |
|
|
| - Benchmark: https://huggingface.co/datasets/OrderLab/distdebug-bench |
| - Code, runner, and evaluator: https://github.com/OrderLab/distdebug-bench |
| - Website and leaderboard: https://distdebug.com |
|
|
| ## Contents |
|
|
| | Path | Description | |
| |---|---| |
| | `data/attempts.parquet` | One row per attempt: configuration, bug, run, the nine accuracy scores, token use and cost, and the diagnosis. | |
| | `runs/<config>.tar.gz` | Every attempt's full run directory, including the trace. | |
| | `configs.json` | Harness, harness version, and model of each configuration. | |
| | `selection.json` | The run IDs of the three scored attempts per bug and configuration. | |
|
|
| | Configuration | Harness | Model | Archive | |
| |---|---|---|---:| |
| | `mini-swe-agent__gpt-6-astra` | Mini-SWE-Agent 2.2.8 | GPT-6 Astra | 14 MB | |
| | `mini-swe-agent__claude-opus-5.5` | Mini-SWE-Agent 2.2.8 | Claude Opus 5.5 | 10 MB | |
| | `mini-swe-agent__gpt-5.5` | Mini-SWE-Agent 2.2.8 | GPT-5.5 | 17 MB | |
| | `mini-swe-agent__gpt-5.4` | Mini-SWE-Agent 2.2.8 | GPT-5.4 | 103 MB | |
| | `mini-swe-agent__gpt-5.3-codex` | Mini-SWE-Agent 2.2.8 | GPT-5.3-Codex | 18 MB | |
| | `mini-swe-agent__mimo-v2.5-pro` | Mini-SWE-Agent 2.2.8 | MiMo v2.5 Pro | 32 MB | |
| | `mini-swe-agent__kimi-k2.6` | Mini-SWE-Agent 2.2.8 | Kimi K2.6 | 30 MB | |
| | `mini-swe-agent__mimo-v2.5` | Mini-SWE-Agent 2.2.8 | MiMo v2.5 | 34 MB | |
| | `mini-swe-agent__deepseek-v4-flash` | Mini-SWE-Agent 2.2.8 | DeepSeek V4 Flash | 26 MB | |
| | `mini-swe-agent__gpt-5.2-codex` | Mini-SWE-Agent 2.2.8 | GPT-5.2-Codex | 13 MB | |
| | `mini-swe-agent__hy3` | Mini-SWE-Agent 2.2.8 | Hunyuan3 | 20 MB | |
| | `mini-swe-agent__gemini-3.5-flash` | Mini-SWE-Agent 2.2.8 | Gemini 3.5 Flash | 29 MB | |
| | `mini-swe-agent__o4-mini` | Mini-SWE-Agent 1.14.4, 2.2.8 | o4-mini | 23 MB | |
| | `codex-cli__gpt-5.2-codex` | Codex CLI 0.128.0 | GPT-5.2-Codex | 7 MB | |
| | `codex-cli__o4-mini` | Codex CLI 0.128.0 | o4-mini | 8 MB | |
| | `opencode__gpt-5.2-codex` | OpenCode 1.4.6 | GPT-5.2-Codex | 12 MB | |
| | `opencode__o4-mini` | OpenCode 1.4.6 | o4-mini | 6 MB | |
|
|
| ### Run directories |
|
|
| Each archive unpacks to `<config>/<bug>/run-NN/`. Run IDs are the original |
| ones. The paper scores the first three attempts per bug with a schema-valid |
| diagnosis, so a gap in the numbering means an earlier attempt ended without |
| one. Every run directory has: |
|
|
| - `diagnosis.json`: the final diagnosis (root-cause blocks and an explanation) |
| - `diagnosis_scores.json`: the ground truth, the diagnosis, and its scores |
| - `session_stats.json`: API calls, tokens, cost, and wall time |
|
|
| The trace depends on the harness: |
|
|
| | Harness | Trace files | |
| |---|---| |
| | Mini-SWE-Agent | `trajectory.json`: the run configuration and every model message, tool call, and tool output | |
| | Codex CLI | `codex_exec.log` (Codex's JSON event stream) and `codex_last_message.json` | |
| | OpenCode | `prompt.txt` and `opencode_cli.log` (OpenCode's JSON event stream) | |
|
|
| ### Attempts table |
|
|
| | Column | Description | |
| |---|---| |
| | `config`, `harness`, `model` | Configuration (archive name), harness, and model | |
| | `bug`, `run`, `attempt` | Benchmark instance, run ID, and position (1–3) among the scored attempts | |
| | `harness_version` | Harness version recorded in the trace | |
| | `workspace` | `anonymized` or `legacy`. Legacy attempts ran before workspace paths were anonymized, so their workspace path contains the bug ID. | |
| | `has_trace` | Whether the run directory contains a trace file | |
| | `num_blocks` | Number of root-cause blocks in the diagnosis | |
| | `file_coverage` … `mention_coverage` | The nine accuracy metrics, in [0, 1] | |
| | `api_calls` … `duration_seconds` | `session_stats.json` fields | |
| | `diagnosis` | The raw `diagnosis.json` text | |
|
|
| ```python |
| from datasets import load_dataset |
| |
| attempts = load_dataset("OrderLab/distdebug-bench-traces", split="train").to_pandas() |
| per_bug = attempts.groupby(["config", "bug"])["block_coverage"].mean() |
| print(per_bug.groupby("config").mean().sort_values(ascending=False)) |
| ``` |
|
|
| Notes: |
|
|
| - Nine attempts submitted a diagnosis with no blocks. They count as zero on |
| every metric, as in the paper. |
| - `mini-swe-agent__o4-mini` `ray-60991/run-01` has no trace file. |
| - Codex CLI records one API call and no cost per run. |
| - Claude Opus 5.5's thinking text is hidden, so `output_reasoning_tokens` is |
| 0. Thinking tokens are billed as output tokens, so `total_tokens` and `cost` |
| are correct. |
|
|
| ## Reproducing the paper |
|
|
| The GitHub repository rescores every diagnosis with the public evaluator, |
| aggregates the scores, and rebuilds Table 1 and Figures 2 and 3. It then |
| checks the results against the paper's: |
|
|
| ```bash |
| git clone https://github.com/OrderLab/distdebug-bench && cd distdebug-bench |
| pip install -e ".[paper]" |
| hf download OrderLab/distdebug-bench-traces --repo-type dataset --local-dir traces |
| python paper/reproduce.py traces |
| ``` |
|
|
| ## Redactions |
|
|
| The traces are otherwise unmodified, except for these replacements: |
|
|
| - Local paths, the machine's host name and user name, a private API endpoint, |
| an internal project name, and one git author line are replaced with the |
| placeholders `/home/user`, `host`, `user`, `<azure-endpoint>`, `harness`, and |
| `user <user@example.com>`. |
| - Encrypted reasoning items returned by the OpenAI Responses API |
| (`encrypted_content`) are set to `null`. They are opaque and cannot be |
| decrypted. |
|
|
| Rescoring the redacted diagnoses reproduces every stored score. |
|
|
| ## License |
|
|
| The diagnoses, scores, and metadata are under the MIT license. Tool outputs |
| in the traces quote source code, git history, and issue text from the 11 |
| upstream systems, which remain under their own licenses. See the |
| [benchmark card](https://huggingface.co/datasets/OrderLab/distdebug-bench) for |
| the list. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @inproceedings{jiang2026distdebugbench, |
| title = {{DistDebug-Bench}: Can {LLM} Agents Diagnose Bugs in Distributed Systems?}, |
| author = {Jiang, Yuxuan and Liu, Chenglin and Rose, John and Miller, Mason and Thavamani, Ruthesh and Wu, Juntao and Huang, Yinwei and Xu, Jiahao and Huang, Peng}, |
| booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track}, |
| year = {2026} |
| } |
| ``` |
|
|
