Hugging Face's logo

Datasets:
twothsnd
/
ReTerm-10K

Tasks:
Text Generation
Modalities:
Tabular
Text
Formats:
json
Languages:
English
Size:
1K - 10K
ArXiv:
Tags:
agents
tool-use
terminal-agent
supervised-fine-tuning
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringlengths
37
37
task_id
stringlengths
20
20
messages
listlengths
17
259
train_from_message_index
int64
9
131
target_attempt_index
int64
2
2
trajectory_sha256
stringlengths
64
64
task_000000_f8baca82:127e83933e579d15
task_000000_f8baca82
[ { "content": "You are a helpful assistant that can interact with a Linux terminal.\n\nGoal: complete the user's terminal task by using the available tools.\n\nImportant rules:\n- Use the bash tool to execute exactly one non-interactive shell command at a time.\n- Directory and environment-variable changes are p...
17
2
127e83933e579d155e957b1574bac5f9572209ecb372342730acf0f82ca15edd
task_000000_f8baca82:287f910f793ff342
task_000000_f8baca82
[{"content":"You are a helpful assistant that can interact with a Linux terminal.\n\nGoal: complete (...TRUNCATED)
29
2
287f910f793ff342046d72d4a9a784a9a680c82f4981f3f31c3c5685072c1d65
task_000000_f8baca82:3dd9365521995a0e
task_000000_f8baca82
[{"content":"You are a helpful assistant that can interact with a Linux terminal.\n\nGoal: complete (...TRUNCATED)
23
2
3dd9365521995a0e53ca294f0df19bf06c873cded5b4cc01b4996f88e4e3ab4f
task_000000_f8baca82:8f281f6442a06f51
task_000000_f8baca82
[{"content":"You are a helpful assistant that can interact with a Linux terminal.\n\nGoal: complete (...TRUNCATED)
21
2
8f281f6442a06f51bc394c0e4003988e4ecde79b1486074ea14ac8cabe7c31f9
task_000004_aeddda76:703199e4a6afc729
task_000004_aeddda76
[{"content":"You are a helpful assistant that can interact with a Linux terminal.\n\nGoal: complete (...TRUNCATED)
41
2
703199e4a6afc729f71d922ffc0caa48e5f59186b0e7340de7856287e63cab5d
task_000004_aeddda76:a55ce61cd1f0e883
task_000004_aeddda76
[{"content":"You are a helpful assistant that can interact with a Linux terminal.\n\nGoal: complete (...TRUNCATED)
25
2
a55ce61cd1f0e883d3f28d5c9c6be00c255122728136cd2ef8870f68bb101a98
task_000004_aeddda76:d27c73b320432696
task_000004_aeddda76
[{"content":"You are a helpful assistant that can interact with a Linux terminal.\n\nGoal: complete (...TRUNCATED)
35
2
d27c73b32043269679548d1e68d5e908d0f31cbfc9d801e07f6997d8bc86ea13
task_000005_ceefb305:081d09e47e9423e6
task_000005_ceefb305
[{"content":"You are a helpful assistant that can interact with a Linux terminal.\n\nGoal: complete (...TRUNCATED)
131
2
081d09e47e9423e6f18b36c71fe784c7b0a9b949d35bc2eaa1eafb99b7f8c83f
task_000005_ceefb305:0bed79a79a157af1
task_000005_ceefb305
[{"content":"You are a helpful assistant that can interact with a Linux terminal.\n\nGoal: complete (...TRUNCATED)
79
2
0bed79a79a157af1cfe9c50ddaca4e2edcb837be52cafd7589b5d8b8dfd9d955
task_000006_93ea43e6:b78d28a7b366291a
task_000006_93ea43e6
[{"content":"You are a helpful assistant that can interact with a Linux terminal.\n\nGoal: complete (...TRUNCATED)
19
2
b78d28a7b366291ac23ade6b577308916c65efb8a70a71a5a24eec01701f1e8f
End of preview. Expand in Data Studio

ReTerm-10K

This dataset provides SFT training data for ReTerm, a self-improving terminal agent that learns to recover from errors to improve task completion, as presented in ReTerm: Recovery Training for Self-Improving Terminal Agents.

Paper · Code · ReTerm-9B model

About ReTerm-10K

ReTerm-10K is a high-quality terminal-task recovery trajectory dataset for SFT cold start. A dedicated data pipeline generates the demonstrations: a base model first executes terminal tasks, and a stronger teacher then continues from the environment states left by failed attempts, using the preceding interaction history to perform recovery.

ReTerm-10K construction: collect failed states, generate teacher recoveries, and filter the continuations

Task verification, format, task-constraint, and efficiency filtering yield 9,712 recovery trajectories. These demonstrations include inspecting failed execution states, repairing existing implementations, and verifying task outcomes, teaching the agent to build on useful progress and resolve the problems that prevented completion.

Dataset Files

The data is provided in JSONL format. Each example contains the original task, the base model's failed interaction history, and the teacher's successful recovery trajectory.

Field Description
id, task_id Trajectory and source-task identifiers
messages Task instruction, failed interaction history, and teacher recovery trajectory in message order
train_from_message_index Starting message index of the teacher recovery trajectory
target_attempt_index Index of the recovery attempt
trajectory_sha256 Source-trajectory integrity hash

During SFT, the original task and failed history provide context, while loss is applied only to assistant tokens generated during the teacher's recovery. Training code and example configurations are available in the ReTerm repository.

Acknowledgements and License

We thank TMax-15K for the task resources. The dataset is released under ODC-BY.
Downloads last month
21
Number of rows:
9,712
Total file size:
3.41 GB

Models trained or fine-tuned on twothsnd/ReTerm-10K

Paper for twothsnd/ReTerm-10K