README.md

metadata

pretty_name: Training Data VID2
language:
  - zh
  - en
task_categories:
  - text-generation
tags:
  - sft
  - tool-use
  - video-evaluation
  - qwen3.5

Training Data VID2

This dataset contains tool-using supervised fine-tuning (SFT) trajectories for the video interval evaluation task in ProRL-Agent-Server. The trajectories were primarily rolled out with Qwen/Qwen3.5-122B-A10B-FP8 and are used to train/evaluate the Qwen3.5-4B SFT experiments.

The source directory in the repository is:

examples/video_eval_interval_multi_nodes/SFT_data_curation/sft_qwen35_122b/

Related code is located at:

examples/video_eval_interval_multi_nodes/SFT_data_curation/
  step_1_convert_qwen35_rollouts_to_sft.py
  data_mixing_strategy_common.py
  data_mixing_strategy_band_wise_pareto.py
  data_mixing_strategy_max_n.py
  temp_running_script/build_reward_band_no_gt_datasets.py
  temp_running_script/build_reward_gte_0p3_no_gt_dataset.py

examples/video_eval_interval_multi_nodes/SFT_training/
  train_hf_sft.py
  run_hf_sft_with_periodic_eval.sh
  training_args.json

Record format

Each .jsonl line is one complete multi-turn SFT trajectory. Important top-level fields include:

• messages: system/user/assistant/tool turns. Assistant turns are the SFT targets; tool turns preserve the tool-grounded trajectory.
• tools and tool_choice: tool schemas and tool-selection configuration.
• chat_template and chat_template_kwargs: prompt formatting information.
• model and decoding fields such as temperature, top_p, top_k, and repetition penalties: rollout provenance, not necessarily training-time generation settings.
• metadata: source run, task/example/sample IDs, parsing status, reward, interval/tag metrics, reference label, parsed prediction, and token counts.

Some messages contain embedded visual payloads, so individual records and files are unusually large. Use streaming line-by-line reads instead of loading an entire JSONL file into memory.

import json

with open(
    "original/0.7-1.0/"
    "qwen35_122b_sft_train_full_n16_tool_sft_high_quality.jsonl"
) as f:
    for line in f:
        row = json.loads(line)
        messages = row["messages"]
        reward = row["metadata"]["reward"]
        # train on / inspect one trajectory at a time

Files included in this Hugging Face dataset

Only the following paths are uploaded:

README.md
original/
  0.3-0.5/
  0.5-0.7/
  0.7-1.0/

The derived JSONL files stored at the local directory root are intentionally not uploaded. This keeps the remote dataset limited to reusable source-level converted trajectories; merged, sampled, deduplicated, and Pareto variants can be regenerated in the ProRL-Agent-Server repository.

original/ reward bands

The reward intervals use an inclusive lower bound:

• original/0.3-0.5/: [0.3, 0.5)
• original/0.5-0.7/: [0.5, 0.7)
• original/0.7-1.0/: [0.7, 1.0]

Each band contains the same source categories when qualifying trajectories are available:

Filename pattern	Meaning
m40_qwen35_122b_sft_train_full_n16_m40_4gpu_max65_*	General training rollouts produced on the m40 environment, with the run-specific n16, four-GPU, and max-65 settings.
qwen35_122b_sft_train_full_n16_*	General training rollouts with up to 16 sampled trajectories per case.
m40_qwen35_122b_sft_difficult_lv2_n32_*	Difficult-level-2 training rollouts with up to 32 samples per case.
qwen35_122b_sft_difficult_lv1_n16_*	Difficult-level-1 training rollouts with up to 16 samples per case.
qwen35_sft_val_n64_2x4_after_random_*	Validation trajectories generated by the n64_2x4_after_random rollout run.
qwen35_sft_gt_guided_difficult_lv1_val_n20_*	Difficult-level-1 validation trajectories rolled out with GT as privileged context.
qwen35_sft_gt_guided_difficult_lv3_n8_*	Difficult-level-3 trajectories rolled out with GT as privileged context.
*.lengths.json	Precomputed Qwen3.5-4B token lengths for the adjacent JSONL file, used for length grouping/bucketing.

All JSONL filenames end in tool_sft_high_quality, meaning that they contain converted tool-trajectory SFT records that passed the applicable conversion and quality checks. The m40 component records the rollout environment; n8, n16, n20, and n32 record the run's sampling count.

Two paths are intentionally present but empty because that reward band had no matching GT-guided rows:

original/0.5-0.7/qwen35_sft_gt_guided_difficult_lv1_val_n20_tool_sft_high_quality.jsonl
original/0.5-0.7/qwen35_sft_gt_guided_difficult_lv3_n8_tool_sft_high_quality.jsonl

Do not interpret these zero-byte placeholders as corrupt uploads.

Regenerating the derived experiment files

The repository's data-curation scripts merge/filter these original sources to produce experiment-ready no-GT-guided files. For example, the local [0.5, 0.7) experiment uses the generated files:

TRAIN_FILE=qwen35_122b_sft_reward_0p5_to_0p7_no_gt_guided_train.jsonl
VALIDATION_FILE=qwen35_122b_sft_reward_0p5_to_0p7_no_gt_guided_val.jsonl
LENGTHS_FILE=qwen35_122b_sft_reward_0p5_to_0p7_no_gt_guided_train.lengths.json

The repository launchers resolve these paths relative to examples/video_eval_interval_multi_nodes/SFT_data_curation/sft_qwen35_122b/. These generated root-level files are not part of this Hugging Face upload.

Caveats

• The dataset contains large embedded media/tool contexts and totals hundreds of GB; clone/download only the files needed for a specific experiment.
• Absolute local paths inside metadata are provenance from the source machine and are not expected to resolve elsewhere.
• Reward-band datasets are trajectory-level datasets. Multiple rows can share the same underlying example/case.
• GT-guided trajectories can leak privileged information into assistant-side reasoning even after the explicit GT prompt block is removed. Regenerate and use the local no_gt_guided derivatives for controlled no-privilege experiments.
• Verify data rights, privacy requirements, and downstream usage constraints before redistributing or using this dataset outside the research project.