Datasets:
messages listlengths 16 26 | tools listlengths 0 0 | task_id stringclasses 6
values | replica_id int8 3 7 | reward float32 1 1 | model stringclasses 1
value | harness stringclasses 1
value | source_dataset stringclasses 1
value | source_revision stringclasses 1
value | hit_labels listlengths 1 1 | trigger_command stringclasses 6
values | recovery_action stringclasses 3
values | trajectory_format stringclasses 1
value | collection_config stringclasses 1
value | trajectory_sha256 stringclasses 6
values |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
[
{
"role": "user",
"content": "You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be given a task description and the output from previously executed commands. Your goal is to solve the task by providing batches of shell commands.\n\nFormat your response as JSON w... | [] | tmax/task_000043_12653d83 | 7 | 1 | DeepSeek-V4-Flash-0731 | Harbor 0.7.0 / Terminus-2 2.0.0 JSON parser | TMaxxx/TMax-15K-Harbor | 48a77eb0b017606c643ed905f96db41672914798 | [
"git_pager_q_exit"
] | cd /home/user/geoparse-repo && git log -p -- parser.py | {"kind": "q", "reference": "trajectory.json:step-5:command-0"} | ATIF raw_content=true linear_history=true; Harbor last-episode export | {"max_input_tokens": 48640, "max_output_tokens": 16384, "max_turns": 200, "model": "openai/dsv4-flash-0731", "production_concurrency": {"final_recovery_total": 16, "initial_requested_total": 32, "worker_slots": 8}, "served_model": "dsv4-flash-0731", "temperature": 1.0, "thinking": {"reasoning_effort": "high", "thinking... | d211acef6584d2fadbb8fd208dcebdad73312dd5f5383d4b467dc3b3c4cf3153 |
[
{
"role": "user",
"content": "You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be given a task description and the output from previously executed commands. Your goal is to solve the task by providing batches of shell commands.\n\nFormat your response as JSON w... | [] | tmax/task_000873_a7f964b2 | 4 | 1 | DeepSeek-V4-Flash-0731 | Harbor 0.7.0 / Terminus-2 2.0.0 JSON parser | TMaxxx/TMax-15K-Harbor | 48a77eb0b017606c643ed905f96db41672914798 | [
"git_pager_q_exit"
] | cd /home/user/pipeline_repo && git show --stat --oneline acf745c && echo '--- initial ---' && git show acf745c | {"kind": "q", "reference": "trajectory.json:step-7:command-0"} | ATIF raw_content=true linear_history=true; Harbor last-episode export | {"max_input_tokens": 48640, "max_output_tokens": 16384, "max_turns": 200, "model": "openai/dsv4-flash-0731", "production_concurrency": {"final_recovery_total": 16, "initial_requested_total": 32, "worker_slots": 8}, "served_model": "dsv4-flash-0731", "temperature": 1.0, "thinking": {"reasoning_effort": "high", "thinking... | d404c6816db0a4eb61016afa177f9d34962af4feec0f0373e036045299a5b76b |
[{"role":"user","content":"You are an AI assistant tasked with solving command-line tasks in a Linux(...TRUNCATED) | [] | tmax/task_002510_12753f45 | 7 | 1 | DeepSeek-V4-Flash-0731 | Harbor 0.7.0 / Terminus-2 2.0.0 JSON parser | TMaxxx/TMax-15K-Harbor | 48a77eb0b017606c643ed905f96db41672914798 | [
"git_pager_q_exit"
] | "cd /home/user/optimization_engine && git show --stat --oneline $(git rev-parse v1.0-good) && git lo(...TRUNCATED) | {"kind": "q", "reference": "trajectory.json:step-7:command-0"} | ATIF raw_content=true linear_history=true; Harbor last-episode export | "{\"max_input_tokens\": 48640, \"max_output_tokens\": 16384, \"max_turns\": 200, \"model\": \"openai(...TRUNCATED) | f29ec04b9aabdb734d50451ad92a964d293db069ccc584774298ba2956562a50 |
[{"role":"user","content":"You are an AI assistant tasked with solving command-line tasks in a Linux(...TRUNCATED) | [] | tmax/task_003575_5cda4281 | 3 | 1 | DeepSeek-V4-Flash-0731 | Harbor 0.7.0 / Terminus-2 2.0.0 JSON parser | TMaxxx/TMax-15K-Harbor | 48a77eb0b017606c643ed905f96db41672914798 | [
"git_pager_q_exit"
] | git log --oneline -- src/pipeline.py __init__.py src/__init__.py | {"kind": "q", "reference": "trajectory.json:step-8:command-0"} | ATIF raw_content=true linear_history=true; Harbor last-episode export | "{\"max_input_tokens\": 48640, \"max_output_tokens\": 16384, \"max_turns\": 200, \"model\": \"openai(...TRUNCATED) | 0c452d5b311298168a27059a2d80b4cb1ee16e77cf48c41bff8442047fed7809 |
[{"role":"user","content":"You are an AI assistant tasked with solving command-line tasks in a Linux(...TRUNCATED) | [] | tmax/task_004516_b90185a3 | 4 | 1 | DeepSeek-V4-Flash-0731 | Harbor 0.7.0 / Terminus-2 2.0.0 JSON parser | TMaxxx/TMax-15K-Harbor | 48a77eb0b017606c643ed905f96db41672914798 | [
"git_pager_q_exit"
] | cd /app/wal_processor && git log -p --all -S 'all_events' -- processor.py | {"kind": "q", "reference": "trajectory.json:step-5:command-0"} | ATIF raw_content=true linear_history=true; Harbor last-episode export | "{\"max_input_tokens\": 48640, \"max_output_tokens\": 16384, \"max_turns\": 200, \"model\": \"openai(...TRUNCATED) | c087e5e16c45c31275f667a22643cc8ffa2e3f9210e8ef910f0085885c7e6cc4 |
[{"role":"user","content":"You are an AI assistant tasked with solving command-line tasks in a Linux(...TRUNCATED) | [] | tmax/task_004715_e1f56e7b | 4 | 1 | DeepSeek-V4-Flash-0731 | Harbor 0.7.0 / Terminus-2 2.0.0 JSON parser | TMaxxx/TMax-15K-Harbor | 48a77eb0b017606c643ed905f96db41672914798 | [
"git_pager_q_exit"
] | git log --oneline --graph --decorate --all -40 | {"kind": "q", "reference": "trajectory.json:step-5:command-0"} | ATIF raw_content=true linear_history=true; Harbor last-episode export | "{\"max_input_tokens\": 48640, \"max_output_tokens\": 16384, \"max_turns\": 200, \"model\": \"openai(...TRUNCATED) | 3460ddf408aace8c5a3d62fd179c3be3800bf930063bd7603832122404f46007 |
DeepSeek V4 Flash TMax Git Pager Recovery
This dataset contains 6 manually audited, SFT-ready terminal-agent trajectories generated by DeepSeek-V4-Flash-0731 in public TMax environments. The primary subset is deliberately narrow: the agent must actually enter a Git pager or foreground TUI, execute a useful recovery action, return to a shell prompt, and finish the task with reward 1.0.
Source and collection
- Environment/task source:
TMaxxx/TMax-15K-Harbor, pinned to48a77eb0b017606c643ed905f96db41672914798. - Canonical TMax documentation and license:
allenai/TMax-15K(ODC-BY) and the TMax repository. - Selection: 94 public tasks whose instructions explicitly combine approximately 200 commits with required Git-history, regression, or bisection investigation.
- Sampling: 752 rollouts were planned across 94 tasks; 744 valid model attempts were retained for screening and 590 received reward 1.0. Coverage was 87 task(s) with 8/8 valid rollouts, 6 task(s) with 7/8 valid rollouts, 1 task(s) with 6/8 valid rollouts.
- Model:
openai/dsv4-flash-0731, served IDdsv4-flash-0731. - Sampling settings: temperature 1.0, top-p 0.95, maximum 200 Terminus-2 turns.
- Thinking request:
extra_body.chat_template_kwargs.thinking=trueandreasoning_effort=high. - Harness: Harbor 0.7.0, Terminus-2 2.0.0, JSON parser.
- Trajectories: Harbor ATIF with
raw_content=trueandlinear_history=true; only reasoning content actually returned by the server is retained.
The frozen ordered task-list SHA256 is 91587344fd673d4213c98134b2649acbf4d43b274d737b4cd5a64b8ae3cdc529. The published data/train.parquet SHA256 is 8497f98dadb69d9401aaa19943585ca497b697b8e6e5a928af92d5234249ee40.
Filtering and manual audit
A deterministic classifier required both pager/TUI-entry evidence and a later executed recovery action. Proactive git --no-pager or GIT_PAGER=cat use without a prior takeover was recorded but excluded. Every deterministic hit was then manually reviewed against its full ATIF trajectory, terminal observations, interactive recording, and verifier result. Failed tasks, unrecovered hangs, missing responses, parser/protocol defects, fabricated output, and infrastructure-contaminated attempts were excluded.
Recovery labels:
git_pager_q_exit: 6
Rejected and raw trajectories remain local and are not published. This creates intentional selection bias: the dataset teaches recovery from a foreground pager/TUI and is not representative of general terminal-agent behavior or overall TMax pass rate.
Files and schema
data/train.parquet: one full last-episode Terminus-2 conversation per approved trajectory.data/index.parquet: task, replica, reward, evidence label, and non-sensitive checksum metadata.manifest/tasks.jsonl: source manifest entries for the approved SFT tasks.manifest/selected-sft-tasks.jsonl: approved task identities, normalized RL keys, instruction/task-tree fingerprints, and retained replicas.manifest/run-config.json: non-sensitive collection settings.manifest/checksums.sha256: published-file checksums.scripts/: deterministic selection and pager filtering code.
Important fields include messages, tools, task_id, replica_id, reward, model, harness, source_dataset, source_revision, hit_labels, trigger_command, recovery_action, trajectory_format, and collection_config.
License, attribution, and responsible use
TMax data is distributed under ODC-BY and should be cited as:
Ivison et al., “TMax: A Simple Recipe for Training Terminal Agents,” 2026.
The TMax source card also notes that parts of its source material were produced with Gemini and remain subject to applicable Google terms. These trajectories are newly generated DeepSeek outputs. DeepSeek's platform terms assign output rights to the user and permit research, derivatives, and training/distillation; this card discloses their AI-generated nature and does not imply DeepSeek endorsement. Users remain responsible for reviewing the source and model terms for their use case.
Privacy and sanitization
Upload staging was scanned fail-closed for private keys, tokens, authorization headers, .env material, internal endpoints/IPs, private mount/home paths, private GitLab URLs, and unnecessary Slurm identifiers. Secrets and internal infrastructure identifiers were removed. Raw trials, debug request logs, tunnel logs, and rejected trajectories are not included.
- Downloads last month
- 48