docs/data_processing.md

Construction, prompts and temporal processing

Annotation lineage

The source QA annotations come from ThinkStream's streaming_rlvr_processed.jsonl and streaming_cot_cold_processed_5_20.jsonl. The construction pipeline selected verifiable multiple-choice, binary and counting answers and converted them into single-question chat records with source-video intervals.

The earlier base pool combined an 8,343-row pass-rate-selected subset and a 23,884-row verifiable subset, deduplicating on question text and ground truth to produce 25,118 rows. The historical pass-rate selector used eight samples per student/teacher and retained teacher score >= student score while excluding cases where both were perfect. The published artifacts are the retained pool; complete historical generation and sampling traces are not included.

legacy_25118 stores that 25,118-row pool together with its explicit cue-instruction teacher prompt. Ignoring the teacher fields yields the corresponding plain OPD QA inputs. This configuration intentionally retains four previously confirmed unreadable intervals and the eight known overlapping-source records.

grpo_25114 removes those four unreadable intervals and retains the direct-answer prompt suffixes actually used by the GRPO comparison. It does not contain cues and does not apply the later eight-row source-overlap exclusion.

current additionally removes eight records tied to seven confirmed training/OVO source-video overlaps. It stores the original student prompt, cue-positive teacher prompt, raw reference, and matched-empty reference. metadata/exclusions.json records the exact stable IDs and reasons, without internal filesystem information.

Release transformations

The publication conversion:

1. Replaces internal absolute video paths with source-relative paths.
2. Renames the original numeric row index to source_row_id and adds a split-aware sample_id; the local conversion tool restores the trainer's row-index alias.
3. Gives all configurations one explicit Arrow schema and removes embedded pandas metadata. No binary video data is embedded.
4. Reconstructs convenience reference fields for historical configurations. It does not claim these references were used in historical training.
5. Replaces email text in two questions' answer choices with [REDACTED_EMAIL] in all corresponding prompts. The source IDs are 9390 and 20692; answer letters and ordering are unchanged. privacy_redacted marks those rows.

Except for these two documented contact-text replacements, original questions, answers, cue text, prompt wrappers, video interval values and row order are preserved. No credentials, machine addresses, private author names, internal logs or source absolute paths are intentionally distributed.

Video intervals and frames

video_start and video_end are seconds relative to the referenced video file. The conversion uses a maximum preceding window of 30 seconds:

start = max(0, query_timestamp - 30)
end   = query_timestamp

A timestamp audit matched all 25,106 current rows to source query timestamps and found their video_end equal to the query time. 445 training rows have nonzero start times. Stored intervals are authoritative for reconstruction; do not replace them with whole-video sampling.

The verified training input preflight used decord on CPU, one decoder thread, and qwen-vl-utils 0.0.14 frame-range and frame-count helpers. Frame indices were generated by uniformly spacing between the decoder-resolved start and end indices and rounding to integer frame indices. The four unreadable intervals failed this input check.

The release does not bake a universal FPS, number of frames or pixel budget into the dataset. Actual frame counts and resizing depend on the model processor and resolved training configuration. Keep those settings with experiment records. Qwen3.5 and Qwen3-VL use different processors and model modes. For CueGate, the teacher positive/reference scoring branches reuse the student's decoded visual payload while changing the text prompt.

Four recent frames at deployment are a separate evaluation protocol. Do not reinterpret every training example as exactly four frames. Frame-index sampling alone is not a certification of strict absolute presentation-timestamp causality.

Visual cues and reference prompts

Existing cues were generated with Qwen3.5-27B, using instructions to point to a time or visual region rather than answer the question. Generation stripped recognized answer-option blocks and used automated leakage checks/retries. The historical stripping did not cover every option layout, and automated acceptance does not prove cue correctness or absence of answer information.

The cue generator passed a source-video URL without applying the student's stored start/end interval. Consequently it could inspect the full source clip, including post-query content. These existing annotations must not be described as prefix-only. A later prefix/full pilot is outside this release; its outputs are not mixed in.

For a row with a cue:

• Student and raw reference: original question and options.
• Teacher positive: original question plus the cue instruction wrapper and actual cue.
• Empty reference: same wrapper, with the cue content empty.

For a fallback row, all three teacher message fields equal the original student prompt, cue_available is false, and cue_text is null. Validation has no cues.

Known data limitations

An existing audit identified 252 current MCQs without an explicit option block. The stable IDs are retained in metadata/known_issues.json; this publication does not invent missing choices or claim to resolve the answer mapping. Known source overlap exclusion is incomplete across benchmarks, and no exhaustive human cue leakage filter has been applied. Training/validation video IDs should be inspected for any additional split-specific use case.