dixiyao/FoT_running_logs

Running Logs for FoT

We release partial running logs of Federation over Text: Insight Sharing for Multi-Agent Reasoning for reference and future research.

The logs are organized into client-side problem and task records and server-side insight libraries and profiling artifacts. Both use benchmark names to organize their contents.

Repository structure

best_run_logs/
├── clients/
│   ├── aime26/
│   │   └── problems/problem_XXXX.json
│   ├── gpqa_diamond/
│   │   └── problems/problem_XXXX.json
│   ├── hle/
│   │   └── problems/problem_XXXX.json
│   ├── livecodebench_v6/
│   │   └── problems/problem_XXXX.json
│   ├── claweval/
│   │   └── problems/<task_id>/trace.jsonl
│   ├── pinchbench_v1/
│   │   ├── metrics_log.json
│   │   └── problems/<task_id>/
│   │       ├── problem.json
│   │       ├── agent_config.json
│   │       └── transcript.jsonl
│   └── pinchbench_v2/
│       ├── metrics_log.json
│       └── problems/<task_id>/
│           ├── problem.json
│           ├── agent_config.json
│           └── transcript.jsonl
└── server/
    └── <benchmark>/
        ├── insight_library.json
        ├── profiling.json
        └── profiling_raw_response.txt

The server/ directory contains the same seven benchmark directories as clients/: aime26, gpqa_diamond, hle, livecodebench_v6, claweval, pinchbench_v1, and pinchbench_v2.

Client-side logs

The following counts describe the problem files or task directories included in this release, rather than the full size of each original benchmark.

Directory under clients/	Included problems/tasks	Record format
aime26	30	One JSON file per problem
gpqa_diamond	198	One JSON file per problem
hle	2,487	One JSON file per problem
livecodebench_v6	1,055	One JSON file per problem
claweval	198	One task directory containing trace.jsonl
pinchbench_v1	23	One task directory containing JSON records and a JSONL transcript
pinchbench_v2	124	One task directory containing JSON records and a JSONL transcript

Problem-level JSON records

For aime26, gpqa_diamond, and livecodebench_v6, each problem_XXXX.json contains the problem text and identifier, a recorded solution and predicted answer, a correctness flag, and execution metadata such as output-token counts and loop counts. Some records also include reasoning-token counts and a worker identifier.

For hle, each problem file stores the problem text and identifier, an insight_book containing named textual entries, an iteration value, a correctness flag, and execution metadata. These files use a different schema from the solution-and-answer records above.

Use the problem_id inside each record when identifying a problem; the numeric filename and the stored identifier need not be the same. Field availability and value types vary across benchmarks and records.

Agent task logs

• ClawEval: trace.jsonl contains one JSON event per line. Events include task metadata, messages, tool interactions, and evaluation summaries. The fields depend on the event type.
• PinchBench v1/v2: problem.json records the task prompt, execution status, grading information, and available task-level metadata. agent_config.json records the agent configuration, including model settings, prompts, and tools. transcript.jsonl stores the message sequence, with one JSON record per line. Each benchmark also has a metrics_log.json containing run-level metadata, summary metrics, and task entries.

Server-side artifacts

Each benchmark directory under server/ contains:

File	Contents
insight_library.json	A mapping from named insight identifiers to textual insights.
profiling.json	Structured profiling information, including insight clusters and relationships; some files also contain raw profiling text.
profiling_raw_response.txt	The saved raw text response associated with profiling.

The matching benchmark names provide a way to navigate between client logs and server artifacts. The server files are organized at the benchmark level, rather than as one file per client problem.

Access and download

Downloading directly from this Hugging Face repository requires approval from the dataset authors. Sign in to Hugging Face, submit an access request on the dataset page, and wait for approval. This repository access process does not restrict redistribution permitted by the dataset license.

After approval, authenticate with a token belonging to your approved account and download the logs:

hf auth login
hf download dixiyao/FoT_running_logs \
  --repo-type dataset \
  --include "best_run_logs/**" \
  --local-dir FoT_running_logs

To download only one benchmark's client logs and server artifacts, narrow the include patterns. For example:

hf download dixiyao/FoT_running_logs \
  --repo-type dataset \
  --include "best_run_logs/clients/aime26/**" "best_run_logs/server/aime26/**" \
  --local-dir FoT_running_logs

These files are released as a collection of JSON, JSONL, and text artifacts. Read JSON files as individual objects and JSONL files one line at a time; the collection does not use a single shared tabular schema.

Final note: permitted use and sensitive information

This dataset is licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0); the full license is included in LICENSE. Noncommercial redistribution of unchanged copies is permitted without additional author approval, provided you give appropriate credit to the dataset authors, link to the dataset and license, and retain the required notices. Commercial use and distribution of adapted material are not permitted under this license. You may make adaptations for your own noncommercial use, but may not share them under this license.

Raw running logs may contain sensitive information, including author identities, server names, hostnames, network addresses, access tokens, or other credentials. Their presence does not authorize access to any account, service, or system. Malicious activity is strictly forbidden, including unauthorized use or testing of exposed credentials, unauthorized access, attacks on servers, exploitation of vulnerabilities, service disruption, impersonation, harassment, or unlawful disclosure of personal information. Any security testing requires separate authorization from the affected system owner.

If you encounter a suspected credential or other sensitive detail, please report it privately to the dataset authors without testing it or publicizing it. This security notice does not add restrictions to the rights granted by CC BY-NC-ND 4.0. The license does not grant rights to access third-party systems or override privacy rights, applicable laws, third-party licenses, or legal exceptions and limitations.