Datasetpublic

zhiyuanhucs/agentic-sft-v4-teacher-v2

Agentic SFT trajectories from DeepSeek-V4-Flash (v2) 1,463 verified-correct agent trajectories selected for supervised fine-tuning of NVIDIA Nemotron-3-Ultra. This is the student-filtered successor to zhiyuanhucs/agentic-sft-v4-teacher-v1. V1 applied teacher-side correctness and trajectory-quality filters. V2 also runs the intended student on the teacher-solved tasks and removes tasks the student already solves, while deterministically retaining 15% of those solved tasks as… See the full description on the dataset page: https://huggingface.co/datasets/zhiyuanhucs/agentic-sft-v4-teacher-v2.

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes34downloads

Dataset Card

Agentic SFT trajectories from DeepSeek-V4-Flash (v2)

1,463 verified-correct agent trajectories selected for supervised fine-tuning of NVIDIA Nemotron-3-Ultra. This is the student-filtered successor to `zhiyuanhucs/agentic-sft-v4-teacher-v1`.

V1 applied teacher-side correctness and trajectory-quality filters. V2 also runs the intended student on the teacher-solved tasks and removes tasks the student already solves, while deterministically retaining 15% of those solved tasks as anchors against distribution shift.

Subsets

Config	Contents	Trajectories	Source-task IDs
`all` (default)	All three sources	1,463	858
`swe`	Both SWE-rebench sources	1,118	659
`swe_rebench`	SWE-rebench (2025-12, Python)	734	422
`swe_rebench_v2`	SWE-rebench V2 (multi-language)	384	237
`tmax`	TMax-15K, TB2.1-matched sampling	345	199

The 858 figure is the sum of task IDs within each source. There are 737 unique raw task strings across the combined files because the two SWE-rebench releases share some identifiers. The files are intentionally kept as used for training rather than silently cross-source deduplicated.

python

from datasets import load_dataset

all_data = load_dataset(
    "zhiyuanhucs/agentic-sft-v4-teacher-v2",
    split="train",
)
swe_only = load_dataset(
    "zhiyuanhucs/agentic-sft-v4-teacher-v2",
    "swe",
    split="train",
)

Composition and filtering

Source	Trajectories	Tasks kept	Solved-task anchors	Student-solved tasks removed	Teacher-unsolved tasks	Median steps	Ran own tests
SWE-rebench	734	422	26	113	290	44	678/734
SWE-rebench V2	384	237	20	132	175	58.5	327/384
TMax-15K	345	199	90	571	230	18	34/345
Total	1,463	858	136	816	695	—	1,039/1,463

Selection is performed in this order:

1.The DeepSeek-V4-Flash teacher run must pass the task's own verifier.
2.The Nemotron student is evaluated with pass@2; student-solved tasks are removed, except for a deterministic 15% task-name-hash anchor sample.
3.Runs over 120 steps, under 2 steps, dominated by repeated commands, or with extreme write-versus-explore behavior are rejected.
4.At most two trajectories per task are retained, preferring runs that execute tests and then shorter runs.

Fields

Each JSONL row is one complete trajectory:

Field	Meaning
`task`	Upstream task identifier
`source`	Internal teacher collection shard
`steps`	Number of assistant turns
`ran_tests`	Whether the teacher explicitly ran a recognized test command
`messages`	Full system/user/assistant/tool conversation, including `reasoning_content` and tool calls

messages is the raw conversation used to build the SFT data. It is not pretokenized: model-specific chat-template rendering and loss masking should be performed downstream.

Upstream lookup

Config	Upstream dataset	Match `task` against
`swe_rebench`	`nebius/SWE-rebench`	`instance_id`
`swe_rebench_v2`	`nebius/SWE-rebench-V2`	`instance_id`
`tmax`	`allenai/TMax-15K`	`task_id`

Caveats

—Correct-by-tests is not equivalent to an ideal patch. The quality filters remove obvious bad demonstrations, not every stylistic defect.
—Some older SWE-rebench task directories were no longer available when the student baseline was run. Those trajectories were retained without student evidence and may contain tasks the student could already solve.
—The two SWE-rebench releases have overlapping task identifiers; no cross-source deduplication was applied.
—This repository contains training trajectories, not a benchmark split. Audit overlap before evaluating on any related repository/task collection.