shahedm2001/cataract_surgery_vlm_eval

Cataract Surgery VLM Evaluation Dataset — Test

Self-contained, flat evaluation set for cataract-surgery video understanding. No external dependencies — every record and its video live in this single folder.

At a glance

Item	Count	Notes
JSONL records	1004	one file = one record = one valid JSON line
Videos (.mp4)	518	flat, prefixed names; ~3.45 GB total
Split	Test only	held-out evaluation, no training overlap
Folder layout	flat	no subfolders; every video field is a relative filename in this folder

Record breakdown (1004)

Task category	question_type values	Records	Track
Visual description (open-ended)	visual_description	293 (162 YouTube + 131 phase)	youtube / phase
Multiple-choice (MCQ)	step_identification, instrument_identification, visual_observation	486 (162 each)	youtube
Phase understanding	boundary_detection 49, temporal_localization 54, timestamp_to_phase 47, contextual_phase_recognition 60	210	phase
Procedural narration	narration	15	youtube

Videos used:

• 293 clip videos (162 YouTube + 131 phase) — shared by description and MCQ where applicable
• 210 phase-task clips (task-specific composites)
• 15 full-length surgeries (uncut YouTube procedures)
• Total unique .mp4 = 518 (YT clips reused by description+MCQ are stored once)

Procedures in Test:

• YouTube: 15 (0xUbMicNy-w, 2PnbLBFKiyQ, 9E_jRdOnuqw, CztPCggNaUA, G1Rpj8EoEGE, IFNTpWIMgsQ, JYiu3W-86zg, M8c_NoP01_A, QJwvHv9pBxU, _QXWa7QaEgk, gT1UMavejVQ, if2P7EPOgsY, vEA2kqxaXjk, wMXS1VGKfNY, zaL9vzyCJSI)
  • Clip counts vary (1–16 per procedure; two procedures have a single clip each)
• Phase: 23 (PH_0003_2933_S2 … PH_0150_2495_S1)
  • 1–13 clips per procedure for description; 0–20 task clips per procedure for phase tasks (4 procedures have description clips only and no task clips)

---

File naming

All names are prefixed by the parent procedure ID so the flat folder has no collisions:

Record type	JSONL	Video
Visual description	<ID>_clip_XX_visual_description.jsonl	<ID>_clip_XX.mp4
MCQ	<ID>_clip_XX_<question_type>.jsonl	<ID>_clip_XX.mp4
Phase task	<PH_ID>_<NNNN>_<question_type>.jsonl	<PH_ID>_<NNNN>_<question_type>.mp4
Narration	<ID>_full_video_narration.jsonl	<ID>_full_video.mp4

Examples:

• 0xUbMicNy-w_clip_03_visual_description.jsonl + 0xUbMicNy-w_clip_03.mp4
• 0xUbMicNy-w_clip_03_step_identification.jsonl (shares the same clip mp4)
• PH_0003_2933_S2_0121_boundary_detection.jsonl + PH_0003_2933_S2_0121_boundary_detection.mp4
• IFNTpWIMgsQ_full_video_narration.jsonl + IFNTpWIMgsQ_full_video.mp4

---

JSON schema

Every file contains a single JSON object (one line). Common envelope:

{
  "record_id": "<matches filename stem>",
  "task_category": "visual_description | mcq | phase | narration",
  "question_type": "<specific type>",
  "reward_type": "llm_judge | deterministic",
  "split": "Test",
  "track": "youtube | phase",
  "video": "<flat .mp4 filename in this folder>",
  "metadata": { ... }
}

1) Visual description

{
  "record_id": "0xUbMicNy-w_clip_01_visual_description",
  "task_category": "visual_description",
  "question_type": "visual_description",
  "reward_type": "llm_judge",
  "video": "0xUbMicNy-w_clip_01.mp4",
  "messages": [
    {"role": "user", "content": [
      {"type": "video", "video": "0xUbMicNy-w_clip_01.mp4"},
      {"type": "text", "text": "Describe what is happening in this cataract surgical video clip."}
    ]},
    {"role": "assistant", "content": "A keratome is used to create a paracentesis incision ..."}
  ],
  "metadata": {
    "parent_video_id": "0xUbMicNy-w",
    "parent_video_title": "CataractCoach™ 2657 ...",
    "parent_video_url": "https://www.youtube.com/watch?v=0xUbMicNy-w",
    "step_number": 1,
    "step_title": "paracentesis creation",
    "clip_duration_seconds": 4,
    "timestamp_start_in_parent": "00:18",
    "timestamp_end_in_parent": "00:22",
    "instruments": ["Keratome"],
    "anatomy": ["cornea", "anterior chamber"]
  }
}

Prompt variants are all "Describe what is happening ..." (YouTube) or "Describe the visually observable surgical activity ..." (phase). Phase prompts explicitly forbid inferring beyond visible evidence.

2) Multiple-choice (YouTube)

{
  "record_id": "0xUbMicNy-w_clip_01_step_identification",
  "task_category": "mcq",
  "question_type": "step_identification",
  "reward_type": "deterministic",
  "video": "0xUbMicNy-w_clip_01.mp4",
  "prompt": [
    {"role": "user", "content": [
      {"type": "video", "video": "0xUbMicNy-w_clip_01.mp4"},
      {"type": "text", "text": "Which step is being performed?\nA) ...\nB) ...\n... Respond ONLY with a JSON object with two keys: \"explanation\" and \"answer\". \"explanation\" is 1-3 sentences ... \"answer\" is the single letter A-D. Example: {\"explanation\": \"...\", \"answer\": \"B\"}"}
    ]}
  ],
  "correct_answer": "B",
  "reference_reasoning": "In the video, the surgeon is seen using ...",
  "metadata": { /* same whitelist as above */ }
}

Three question_type values cycle per clip:

• step_identification — which surgical step is shown
• instrument_identification — which instrument is primarily used
• visual_observation — which anatomical / visual cue is present

3) Phase understanding (phase procedures)

All four tasks share the same ontology (13 phases) listed in every prompt:

P01 Incision · P02 Viscoelastic · P03 Capsulorhexis · P04 Hydrodissection
P05 Phacoemulsification · P06 Irrigation-Aspiration · P07 Capsule Polishing
P08 Lens Implantation · P09 Lens Positioning · P10 Viscoelastic Suction
P11 Anterior Chamber Flushing · P12 Tonifying-Antibiotics · P13 Idle

Prompt instructs: Respond with a single JSON object with exactly two keys: "explanation" and "answer".

Expected answer per task:

question_type	correct_answer shape	Example answer
boundary_detection	{"timestamp": float} clip-local seconds	{"timestamp": 7.7}
temporal_localization	{"start": float, "end": float} clip-local	{"start": 9.4, "end": 13.8}
timestamp_to_phase	{"phase_id": "P01"-"P13", "phase_name": "..."}	{"phase_id": "P09", "phase_name": "Lens Positioning"}
contextual_phase_recognition	{"phase_id": "P01"-"P13", "phase_name": "..."}	{"phase_id": "P03", "phase_name": "Capsulorhexis"}

Record envelope matches MCQ (prompt + correct_answer + reference_reasoning + sanitized metadata).

4) Procedural narration (full video)

{
  "record_id": "0xUbMicNy-w_full_video_narration",
  "task_category": "narration",
  "question_type": "narration",
  "reward_type": "llm_judge",
  "video": "0xUbMicNy-w_full_video.mp4",
  "messages": [
    {"role": "user", "content": [
      {"type": "video", "video": "0xUbMicNy-w_full_video.mp4"},
      {"type": "text", "text": "Provide a detailed timestamped narration of this surgical procedure."}
    ]},
    {"role": "assistant", "content": "[00:18 - 00:22] paracentesis creation: A keratome is used ...\n[00:22 - 00:27] anterior capsule staining: Trypan blue dye is injected ..."}
  ],
  "metadata": {
    "parent_video_id": "0xUbMicNy-w",
    "parent_video_title": "CataractCoach™ 2657 ...",
    "n_steps": 15,
    "step_titles": ["paracentesis creation", "anterior capsule staining", "..."]
  }
}

Reference narration is a flowing, chronological concatenation of step descriptions with bracketed intervals.

---

Rewards & evaluation

Two scoring families. Deterministic tasks are 0/1 or continuous in [0,1] without any model judge; open-ended tasks use an LLM judge.

Deterministic rewards

MCQ and phase identification (step/instrument/visual_observation, timestamp_to_phase, contextual_phase_recognition):

R_task = 1.0  if normalize(pred_answer) == gold_answer else 0.0
# pred parsed from JSON key "answer" (letter A-D or phase_id "P01"-"P13", case-insensitive)
# Strict JSON required; fallback regex extraction allowed but format bonus withheld

Boundary detection (timestamp, tau = 1.5 s, exponential decay):

R_task = exp(-abs(t_pred - t_gt) / 1.5)
# 0.0 s error -> 1.000,  0.5 s -> 0.717,  1.5 s -> 0.368,  large error -> ~0

Temporal localization (start, end, interval IoU):

inter = max(0, min(pred_end, gt_end) - max(pred_start, gt_start))
union = (pred_end - pred_start) + (gt_end - gt_start) - inter
R_task = inter / union   # 1.0 perfect overlap, 0.0 no overlap

Format bonus (phase tasks only, decoupled):

R_fmt  = 1.0 if output is strict valid JSON with exactly {explanation, answer} else 0.0
R_total = R_task + 0.05 * R_fmt
# e.g. correct+valid 1.05, correct+malformed 1.00, wrong+valid 0.05, wrong+malformed 0.00

LLM-judge rewards

Scores are integers 0–5. Normalized score = score / 5.

Visual description — single 0–5 against the reference description:

• 5 Excellent · 4 Good · 3 Moderate · 2 Poor · 1 Very Poor · 0 Irrelevant/Contradictory
• Judged on: surgical actions/maneuvers, instruments, anatomy, factuality (no hallucinations)

Procedural narration — five 0–5 dimensions, each judged against the reference narration:

Dimension	What is checked
step_coverage	every major step/phase from the reference is mentioned
chronological_accuracy	narrated order matches the reference order
visual_technical_accuracy	instruments / tissue interactions match the reference substance
narrative_flow	single fluid, well-transitioned narration vs. disjointed list
overall_score	holistic 0–5 (factual/chronological correctness weighted over style)

Be strict: generic boilerplate, hallucinated steps/instruments, or a fluent but factually disconnected narration scores low on the factual dimensions regardless of style.

---

Metadata fields

All metadata values are allowlisted to stay HuggingFace-safe (no local paths):

• YouTube-derived: parent_video_id, parent_video_title, parent_video_url, step_number, step_title, clip_duration_seconds, timestamp_start_in_parent, timestamp_end_in_parent, instruments[], anatomy[], transcript_context, visual_description, clip_id
• Phase-derived: phase_id, phase_name, site (S1/S2), subclip_order, phase_occurrence, previous_phase, next_phase, context_bucket, phase_clip_id
• Narration-derived: n_steps, step_titles[]

Fields intentionally excluded: any local filesystem paths, pipeline internals, or teacher/critic labels.

---

Usage

import json, pathlib

root = pathlib.Path("evaluation_dataset")
# Iterate records
for jf in sorted(root.glob("*.jsonl")):
    rec = json.loads(jf.read_text())
    video_path = root / rec["video"]  # resolves inside the same folder
    if rec["task_category"] == "visual_description":
        messages = rec["messages"]  # user (video+text), assistant (reference)
    elif rec["task_category"] in ("mcq", "phase"):
        prompt = rec["prompt"]       # user (video+text)
        gold   = rec["correct_answer"]
    elif rec["task_category"] == "narration":
        messages = rec["messages"]

# Quick verification: 1004 one-line records, all videos resolve
assert len(list(root.glob("*.jsonl"))) == 1004
assert all(len(open(p).read().strip().splitlines()) == 1 for p in root.glob("*.jsonl"))

For VLM inference, fill the video placeholder with frames sampled from the referenced .mp4 and feed the messages / prompt text as the user query.

---

Provenance

• YouTube corpus: public cataract surgery videos (clip segments with step/instrument annotations + full videos with chronological narrations).
• Phase corpus: cataract phase recognition procedures annotated with the 13-phase ontology; task clips are short composites targeting the four phase-understanding skills above.
• This folder contains only the Test split (held-out). Training/validation splits are not included.

License / citation

Videos include third-party YouTube content — respect the original video licenses when redistributing model outputs or excerpts. If you use this evaluation set, please cite the original data sources and note the reward definitions above.