Dataset Viewer
Auto-converted to Parquet Duplicate
trace_id
stringlengths
36
36
timestamp
stringlengths
32
32
model
stringclasses
1 value
status
stringclasses
2 values
prompt_tokens
int64
14
67.3k
⌀
completion_tokens
int64
1
4.22k
⌀
total_tokens
int64
15
71.1k
⌀
cost
float64
0
0
⌀
request
unknown
response
unknown
cc0bb2b9-9bfc-493c-99d8-cc726902b199
2026-10-28T23:02:36.676894+00:00
google/gemma-4-31B-it
success
14
1
15
null
{ "messages": [ { "role": "user", "content": "hi" } ], "model": "google/gemma-4-31B-it" }
{ "choices": [ "Choices(finish_reason='length', index=0, message=Message(content='Hello', role='assistant', tool_calls=None, function_call=None, provider_specific_fields={'reasoning': None, 'refusal': None}), provider_specific_fields={'token_ids': None, 'stop_reason': None, 'routed_experts': None})" ], "id": ...
64cc6387-51bd-4108-b8bf-37bc4abfda59
2026-10-28T23:02:48.230310+00:00
google/gemma-4-31B-it
success
2,581
32
2,613
null
{ "messages": [ { "role": "user", "content": "Resolve the given issue by editing the repository files directly on a remote machine.\n\nRepository directory on the remote machine: /testbed\n\n## Issue to resolve:\nA direct approach to ITRS to Observed transformations that stays within the ITRS.\n<!-- T...
{ "choices": [ "Choices(finish_reason='tool_calls', index=0, message=Message(content=None, role='assistant', tool_calls=[ChatCompletionMessageToolCall(function=Function(arguments='{\"command\": \"find /testbed -name \\\\\"*itrs*\\\\\" \\\\\"*/altaz*\\\\\" \\\\\"*/hadec*\\\\\"\"}', name='bash'), id='chatcmpl-tool-...
6787533c-b372-467f-852b-f5f33522615d
2026-10-28T23:02:49.713768+00:00
google/gemma-4-31B-it
success
2,662
38
2,700
null
{ "messages": [ { "role": "user", "content": "Resolve the given issue by editing the repository files directly on a remote machine.\n\nRepository directory on the remote machine: /testbed\n\n## Issue to resolve:\nA direct approach to ITRS to Observed transformations that stays within the ITRS.\n<!-- T...
{ "choices": [ "Choices(finish_reason='tool_calls', index=0, message=Message(content=None, role='assistant', tool_calls=[ChatCompletionMessageToolCall(function=Function(arguments='{\"command\": \"find /testbed -name \\\\\"*itrs*\\\\\" -o -name \\\\\"*altaz*\\\\\" -o -name \\\\\"*hadec*\\\\\"\"}', name='bash'), id...
f798fd89-3941-445a-a169-a6c3f01619c6
2026-10-28T23:02:50.519883+00:00
google/gemma-4-31B-it
success
2,843
29
2,872
null
{ "messages": [ { "role": "user", "content": "Resolve the given issue by editing the repository files directly on a remote machine.\n\nRepository directory on the remote machine: /testbed\n\n## Issue to resolve:\nA direct approach to ITRS to Observed transformations that stays within the ITRS.\n<!-- T...
{ "choices": [ "Choices(finish_reason='tool_calls', index=0, message=Message(content=None, role='assistant', tool_calls=[ChatCompletionMessageToolCall(function=Function(arguments='{\"path\": \"/testbed/astropy/coordinates/builtin_frames/itrs.py\"}', name='view'), id='chatcmpl-tool-bc0a2c62586589bc', type='functio...
40ed0c46-ee8e-4854-9c75-fc322211b35b
2026-10-28T23:02:53.189532+00:00
google/gemma-4-31B-it
success
3,467
52
3,519
null
{ "messages": [ { "role": "user", "content": "Resolve the given issue by editing the repository files directly on a remote machine.\n\nRepository directory on the remote machine: /testbed\n\n## Issue to resolve:\nA direct approach to ITRS to Observed transformations that stays within the ITRS.\n<!-- T...
{ "choices": [ "Choices(finish_reason='tool_calls', index=0, message=Message(content=None, role='assistant', tool_calls=[ChatCompletionMessageToolCall(function=Function(arguments='{\"command\": \"grep -r \\\\\"ITRS\\\\\" /testbed/astropy/coordinates/intermediate_rotation_transforms.py || grep -r \\\\\"ITRS\\\\\" ...
41d72570-556d-41b2-a38a-15102f1cf87c
2026-10-28T23:02:54.971061+00:00
google/gemma-4-31B-it
success
6,662
32
6,694
null
{ "messages": [ { "role": "user", "content": "Resolve the given issue by editing the repository files directly on a remote machine.\n\nRepository directory on the remote machine: /testbed\n\n## Issue to resolve:\nA direct approach to ITRS to Observed transformations that stays within the ITRS.\n<!-- T...
{ "choices": [ "Choices(finish_reason='tool_calls', index=0, message=Message(content=None, role='assistant', tool_calls=[ChatCompletionMessageToolCall(function=Function(arguments='{\"path\": \"/testbed/astropy/coordinates/builtin_frames/intermediate_rotation_transforms.py\"}', name='view'), id='chatcmpl-tool-aae4...
9c08fd91-20c7-4fdc-8e2b-5edade6a2548
2026-10-28T23:02:57.596685+00:00
google/gemma-4-31B-it
success
10,910
29
10,939
null
{ "messages": [ { "role": "user", "content": "Resolve the given issue by editing the repository files directly on a remote machine.\n\nRepository directory on the remote machine: /testbed\n\n## Issue to resolve:\nA direct approach to ITRS to Observed transformations that stays within the ITRS.\n<!-- T...
{ "choices": [ "Choices(finish_reason='tool_calls', index=0, message=Message(content=None, role='assistant', tool_calls=[ChatCompletionMessageToolCall(function=Function(arguments='{\"path\": \"/testbed/astropy/coordinates/builtin_frames/altaz.py\"}', name='view'), id='chatcmpl-tool-a36bdd48f9611783', type='functi...
939f24e8-6c5d-4013-ae51-1718d8e59323
2026-10-28T23:02:59.016458+00:00
google/gemma-4-31B-it
success
13,073
34
13,107
null
{ "messages": [ { "role": "user", "content": "Resolve the given issue by editing the repository files directly on a remote machine.\n\nRepository directory on the remote machine: /testbed\n\n## Issue to resolve:\nA direct approach to ITRS to Observed transformations that stays within the ITRS.\n<!-- T...
{ "choices": [ "Choices(finish_reason='tool_calls', index=0, message=Message(content=None, role='assistant', tool_calls=[ChatCompletionMessageToolCall(function=Function(arguments='{\"command\": \"ls /testbed/astropy/coordinates/builtin_frames/icrs_observed_transforms.py\"}', name='bash'), id='chatcmpl-tool-bd54d3...
2ad6e6dc-27d6-4408-b03b-382c869f0743
2026-10-28T23:03:00.481324+00:00
google/gemma-4-31B-it
success
13,150
33
13,183
null
{"messages":[{"role":"user","content":"Resolve the given issue by editing the repository files direc(...TRUNCATED)
{"choices":["Choices(finish_reason='tool_calls', index=0, message=Message(content=None, role='assist(...TRUNCATED)
50b19738-493c-4337-904d-169175f75293
2026-10-28T23:03:10.654868+00:00
google/gemma-4-31B-it
success
15,252
359
15,611
null
{"messages":[{"role":"user","content":"Resolve the given issue by editing the repository files direc(...TRUNCATED)
{"choices":["Choices(finish_reason='tool_calls', index=0, message=Message(content=None, role='assist(...TRUNCATED)
End of preview. Expand in Data Studio

Agent Traces for Position-Independent Caching

These are the agent trajectories we collected ourselves for our study of position-independent KV caching (PIC) in agentic workloads. Every LLM request is logged with the exact messages the agent sent and the raw response. That makes the traces suitable for measuring how much input content repeats within and across sessions. The public datasets the paper also analyzes (SWE-smith, OpenHands, SWE-agent, NNetNav, ITBench, tau2, Exgentic and others) are not copied here. Please get them from their original sources.

Contents

Folder Workload Model Tasks Sessions LLM calls
swarm-swebench-verified/gemma-4-31B-it Coding agent swarm, 9 agents per task Gemma-4-31B-it 10 90 3,981
swarm-swebench-verified/MiniMax-M2.7 Coding agent swarm, 9 agents per task MiniMax-M2.7 10 90 7,977
spider2-dbt/gemma-4-31B-it Database interaction (dbt projects) Gemma-4-31B-it 68 68 1,276

swarm-swebench-verified

Ten SWE-bench Verified tasks, each solved concurrently by nine agents with distinct software-factory roles (re:factory). All nine agents of a task share one repository and one upstream vLLM 0.23.0 server without PIC. Gemma-4-31B-it was served on two GPUs and MiniMax-M2.7 on four. The agents use the Exgentic tool_calling agent on SWE-bench Verified with temperature 1.0, top-p 0.95 and at most 16,384 output tokens per request.

Layout: <model>/<instance_id>/s<agent>/, where each agent directory holds:

  • sessions/<id>/framework/litellm/trace.jsonl: one line per LLM call, with the full request messages, the response and token counts.
  • sessions/<id>/trajectory.jsonl: the agent's step-by-step trajectory.
  • sessions/<id>/session.log: the harness log.
  • sessions/<id>/benchmark/report.json: the SWE-bench grading report (missing when the agent ended without a final patch).
  • sessions/<id>/config.json, run/config.json, results.json: the run configuration and summary.

An agent whose prompt outgrew the context window ended without a report. This happened to 29 of the 90 MiniMax agents and 2 of the 90 Gemma agents. Two further Gemma agents stopped when a reproduction script exceeded the sandbox command limit. These agents count as unresolved.

spider2-dbt

Spider2-DBT tasks run with the official Spider-Agent (methods/spider-agent-dbt) on Gemma-4-31B-it, one run per task. Each task directory holds:

  • calls.jsonl: one line per LLM call, with the exact messages sent and the raw response.
  • result.json: the agent's final answer.

eval.log is the output of the official evaluation suite (8 of 68 tasks correct). The harness windows its history and re-renders past turns, so no single call holds the whole run. To rebuild one trajectory, take the system prompt from the first call and then the newest user message and the model output of each call in order.

Intended use

Studying KV-cache reuse, prompt caching and serving efficiency for multi-turn agents, and analyzing agent behavior.

Downloads last month
83