studyOverflow's picture
fix: count annotator history from both HF snapshot + local file to survive restarts
bffee2e verified
"""
MBench Annotation Space — 单视频标注 + MBench-V Pairwise + MBench-A Pairwise
功能:
- Tab 1 (单视频标注): "该视频是否出现了记忆问题?" (MBench-V)
- Tab 2 (MBench-V Pairwise): 同一 prompt 下两个 T2V 模型视频并排 (MBench-V)
- Tab 3 (MBench-A Pairwise): 世界模型 401f 视频对比,4子集×多维度 (MBench-A)
技术栈:
- Gradio 5.9.1 + FastAPI 视频代理
- HuggingFace CommitScheduler 自动推送标注结果
- 数据来源: studyOverflow/TempMemoryData
部署:
直接替换 HuggingFace Space 的 app.py 即可。
"""
from __future__ import annotations
import json
import os
import random
import threading
import time
import uuid
from collections import defaultdict
from pathlib import Path
from typing import Any
import gradio as gr
from huggingface_hub import CommitScheduler, HfApi, hf_hub_download, hf_hub_url
# ---------------------------------------------------------------------------
# Config
# ---------------------------------------------------------------------------
DATASET_REPO = "studyOverflow/TempMemoryData"
MERGED_JSON_PATH = "MBench-V/merged.json"
MODELS: list[str] = [
"causal_forcing",
"self_forcing",
"cosmos",
"helios",
"longlive",
"memflow",
"longcat",
"skyreels",
]
HF_TOKEN = os.environ.get("HF_TOKEN")
ANN_DIR = Path("annotations_local")
ANN_DIR.mkdir(exist_ok=True)
PROCESS_ID = uuid.uuid4().hex[:8]
# Separate files for annotation types
ANN_FILE_BINARY = ANN_DIR / f"ann_binary_{PROCESS_ID}.jsonl"
ANN_FILE_PAIRWISE = ANN_DIR / f"ann_pairwise_{PROCESS_ID}.jsonl"
ANN_FILE_MBENCH_A = ANN_DIR / f"ann_mbench_a_{PROCESS_ID}.jsonl"
COMMIT_INTERVAL_MIN = 5
PENDING_TIMEOUT_SEC = 30 * 60
# MBench-V Pairwise config
PAIRWISE_DIMENSIONS = [
("entity", "实体一致性", "人物/物体离开画面再回来后,哪个视频中实体外观更一致?"),
("physical", "物理合理性", "哪个视频中的物理过程(水流/碰撞/变形等)更合理自然?"),
("prompt", "Prompt 忠实度", "哪个视频的内容更符合下方的文字描述?"),
]
PAIRWISE_SAMPLES_PER_PAIR = 30
# ---------------------------------------------------------------------------
# MBench-A Config
# ---------------------------------------------------------------------------
MBENCH_A_MODELS: list[str] = [
"hy_worldplay",
"infinite_world",
"lingbot_world",
"matrix_game_2",
"matrix_game_3",
"yume",
]
MBENCH_A_ANNOTATORS_PER_TASK = 3
MBENCH_A_CATEGORY_MAP = {
"environment": "Spatial_401f",
"object": "Spatial_401f",
"human": "Human_401f",
"causal": "Casual_401f",
}
MBENCH_A_GT_CATEGORY_MAP = {
"environment": "Spatial",
"object": "Spatial",
"human": "Human",
"causal": "Casual",
}
# ---------------------------------------------------------------------------
# Load MBench-V merged.json
# ---------------------------------------------------------------------------
def _load_merged() -> list[dict[str, Any]]:
try:
local = hf_hub_download(
repo_id=DATASET_REPO,
filename=MERGED_JSON_PATH,
repo_type="dataset",
token=HF_TOKEN,
)
with open(local, encoding="utf-8") as f:
return json.load(f)
except Exception as e:
print(f"[mbench-ann] WARNING: Failed to load MBench-V data: {e}")
return []
TASKS: list[dict[str, Any]] = _load_merged()
TASK_BY_ID: dict[str, dict[str, Any]] = {t["task_id"]: t for t in TASKS}
# ---------------------------------------------------------------------------
# Load MBench-A task pool
# ---------------------------------------------------------------------------
def _load_mbench_a_pool() -> dict[str, Any]:
"""Load MBench-A task pool from local file or HF."""
local_path = Path(__file__).parent / "sampling" / "task_pool.json"
if local_path.exists():
with open(local_path, encoding="utf-8") as f:
return json.load(f)
# Fallback: try HF
try:
local = hf_hub_download(
repo_id=DATASET_REPO,
filename="MBench-A/task_pool.json",
repo_type="dataset",
token=HF_TOKEN,
)
with open(local, encoding="utf-8") as f:
return json.load(f)
except Exception as e:
print(f"[mbench-ann] WARNING: Failed to load MBench-A task pool: {e}")
return {"tasks": [], "quality_control_tasks": [], "metadata": {}}
MBENCH_A_POOL = _load_mbench_a_pool()
MBENCH_A_TASKS: list[dict] = MBENCH_A_POOL.get("tasks", []) + MBENCH_A_POOL.get("quality_control_tasks", [])
MBENCH_A_TASK_BY_ID: dict[str, dict] = {t["task_id"]: t for t in MBENCH_A_TASKS}
# ---------------------------------------------------------------------------
# MBench-V Pool setup
# ---------------------------------------------------------------------------
BINARY_POOL: list[tuple[str, str]] = [(m, t["task_id"]) for m in MODELS for t in TASKS]
BINARY_POOL_SET: set[tuple[str, str]] = set(BINARY_POOL)
def _build_pairwise_pool() -> list[tuple[str, str, str, str]]:
pool = []
task_ids = [t["task_id"] for t in TASKS[:PAIRWISE_SAMPLES_PER_PAIR]]
for tid in task_ids:
for i, m_a in enumerate(MODELS):
for m_b in MODELS[i+1:]:
for dim_key, _, _ in PAIRWISE_DIMENSIONS:
pool.append((tid, m_a, m_b, dim_key))
return pool
PAIRWISE_POOL: list[tuple[str, str, str, str]] = _build_pairwise_pool()
PAIRWISE_POOL_SET: set[tuple[str, str, str, str]] = set(PAIRWISE_POOL)
print(f"[mbench-ann] MBench-V: {len(TASKS)} tasks × {len(MODELS)} models")
print(f"[mbench-ann] MBench-V binary pool: {len(BINARY_POOL)}, pairwise pool: {len(PAIRWISE_POOL)}")
print(f"[mbench-ann] MBench-A: {len(MBENCH_A_TASKS)} tasks, {len(MBENCH_A_POOL.get('metadata', {}))} metadata")
# ---------------------------------------------------------------------------
# Video URL helpers
# ---------------------------------------------------------------------------
def _video_url(model: str, task_id: str) -> str:
return f"/video/{model}/{task_id}.mp4"
def _hf_video_url(model: str, task_id: str) -> str:
return hf_hub_url(
DATASET_REPO,
filename=f"MBench-V/{model}/videos/{task_id}.mp4",
repo_type="dataset",
)
def _mbench_a_video_proxy_url(model: str, subset: str, sample_id: str) -> str:
"""Build local proxy URL for MBench-A video."""
category = MBENCH_A_CATEGORY_MAP[subset]
return f"/video_a/{model}/{category}/{sample_id}/left_then_right.mp4"
def _mbench_a_hf_video_url(model: str, category: str, sample_id: str) -> str:
"""Build HF upstream URL for MBench-A video."""
return hf_hub_url(
DATASET_REPO,
filename=f"MBench-A/{model}/{category}/{sample_id}/left_then_right.mp4",
repo_type="dataset",
)
def _mbench_a_asset_hf_url(path: str) -> str:
"""Build HF URL for MBench-A assets."""
return hf_hub_url(
DATASET_REPO,
filename=f"MBench-A/assets/{path}",
repo_type="dataset",
)
def _extract_prompt(task: dict[str, Any]) -> str:
gp = task.get("generation_prompts") or {}
prompts = gp.get("prompts") or {}
for level in ("level_3", "level_4", "level_2", "level_1"):
val = prompts.get(level)
if isinstance(val, list) and val:
n = len(val)
return "\n\n".join(f"— 第 {i}/{n} 段 —\n{seg}" for i, seg in enumerate(val, 1))
if isinstance(val, str) and val:
return val
return "(no prompt found)"
def _render_video_html(url: str) -> str:
return (
f'<video controls autoplay muted loop playsinline width="100%" '
f'style="max-height:400px;object-fit:contain" src="{url}">'
f'您的浏览器不支持 HTML5 视频。</video>'
)
# ---------------------------------------------------------------------------
# MBench-A: Auxiliary info rendering
# ---------------------------------------------------------------------------
def _render_mbench_a_aux(task: dict) -> str:
"""Render auxiliary HTML info based on task subset."""
subset = task["subset"]
# Use CSS class for guaranteed visibility (Gradio themes can override inline styles)
box = 'class="aux-info-box"'
# Camera motion info (shown for ALL subsets)
motion = task.get("camera_motion", "left_then_right")
motion_desc = task.get("camera_motion_description", motion)
gif_url = _mbench_a_asset_hf_url(f"camera_diagrams/{motion}.gif")
camera_html = (
f'<div style="flex:0 0 200px">'
f'<p><b>🎬 预期相机运动</b></p>'
f'<p style="margin:0 0 8px">{motion_desc}</p>'
f'<img src="{gif_url}" style="width:180px">'
f'</div>'
)
# Caption (shown for ALL subsets now)
caption = task.get("caption", "")
caption_html = ""
if caption:
caption_html = (
f'<div style="flex:1;min-width:250px">'
f'<p><b>📝 场景描述</b></p>'
f'<p style="font-size:14px;line-height:1.5">{caption}</p>'
f'</div>'
)
if subset == "object":
sample_id = task["sample_id"]
mask_url = _mbench_a_asset_hf_url(f"mask_viz/{sample_id}.png")
return (
f'<div {box}>'
f'<p><b>🎯 请关注画面中被标注(高亮)的物体</b></p>'
f'<div style="display:flex;gap:16px;flex-wrap:wrap;align-items:flex-start;margin-top:8px">'
f'<div style="flex:1;min-width:300px">'
f'<img src="{mask_url}" style="max-width:100%;max-height:280px">'
f'</div>'
f'{camera_html}'
f'{caption_html}'
f'</div></div>'
)
elif subset == "causal":
return (
f'<div {box}>'
f'<div style="display:flex;gap:16px;flex-wrap:wrap;align-items:flex-start">'
f'{camera_html}'
f'{caption_html}'
f'</div></div>'
)
elif subset == "human":
return (
f'<div {box}>'
f'<p><b>👤 请关注视频中的人物</b>:观察人物离开画面再回来后,面部和外观是否保持一致。</p>'
f'<div style="display:flex;gap:16px;flex-wrap:wrap;align-items:flex-start;margin-top:8px">'
f'{camera_html}'
f'{caption_html}'
f'</div></div>'
)
else: # environment
return (
f'<div {box}>'
f'<p><b>🏞️ 请关注整体场景</b>:观察相机转回来后,场景的布局、风格、光照是否保持一致。</p>'
f'<div style="display:flex;gap:16px;flex-wrap:wrap;align-items:flex-start;margin-top:8px">'
f'{camera_html}'
f'{caption_html}'
f'</div></div>'
)
return (
f'<div {box}>'
f'<div style="display:flex;gap:16px;flex-wrap:wrap;align-items:flex-start">'
f'<div style="flex:1;min-width:250px">'
f'<p><b>🏞️ 请关注整体场景</b>:观察相机转回来后,场景的布局、风格、光照是否保持一致。</p>'
f'</div>'
f'{camera_html}'
f'</div></div>'
)
# ---------------------------------------------------------------------------
# CommitScheduler
# ---------------------------------------------------------------------------
scheduler: CommitScheduler | None = None
if HF_TOKEN:
scheduler = CommitScheduler(
repo_id=DATASET_REPO,
repo_type="dataset",
folder_path=str(ANN_DIR),
path_in_repo="annotations",
every=COMMIT_INTERVAL_MIN,
token=HF_TOKEN,
private=False,
squash_history=False,
)
# ---------------------------------------------------------------------------
# Historical annotations
# ---------------------------------------------------------------------------
def _fetch_remote_annotations() -> list[dict[str, Any]]:
records: list[dict[str, Any]] = []
try:
api = HfApi(token=HF_TOKEN)
files = api.list_repo_files(repo_id=DATASET_REPO, repo_type="dataset")
except Exception:
return records
jsonls = [p for p in files if p.startswith("annotations/") and p.endswith(".jsonl")]
for path in jsonls:
try:
local = hf_hub_download(repo_id=DATASET_REPO, filename=path, repo_type="dataset", token=HF_TOKEN)
with open(local, encoding="utf-8") as f:
for line in f:
line = line.strip()
if line:
try:
records.append(json.loads(line))
except Exception:
pass
except Exception:
pass
return records
HISTORICAL = _fetch_remote_annotations()
# ---------------------------------------------------------------------------
# Shared state
# ---------------------------------------------------------------------------
STATE_LOCK = threading.Lock()
# Binary state
BINARY_SUBMITTED: set[tuple[str, str]] = {
(r["model"], r["task_id"]) for r in HISTORICAL
if r.get("type", "binary") == "binary" and "model" in r and "task_id" in r
and (r["model"], r["task_id"]) in BINARY_POOL_SET
}
BINARY_PENDING: dict[tuple[str, str], tuple[str, float]] = {}
# MBench-V Pairwise state
PAIRWISE_SUBMITTED: set[tuple[str, str, str, str]] = {
(r["task_id"], r["model_a"], r["model_b"], r["dimension"])
for r in HISTORICAL
if r.get("type") == "pairwise"
and all(k in r for k in ("task_id", "model_a", "model_b", "dimension"))
}
PAIRWISE_PENDING: dict[tuple[str, str, str, str], tuple[str, float]] = {}
# MBench-A state: task_id -> list of annotators who completed it
MBENCH_A_COMPLETED: dict[str, list[str]] = defaultdict(list)
for r in HISTORICAL:
if r.get("type") == "pairwise_mbench_a" and "task_id" in r and "annotator" in r:
tid = r["task_id"]
# Handle old format where task_id might be stored differently
if tid in MBENCH_A_TASK_BY_ID:
MBENCH_A_COMPLETED[tid].append(r["annotator"])
MBENCH_A_PENDING: dict[str, tuple[str, float]] = {}
print(f"[mbench-ann] binary submitted: {len(BINARY_SUBMITTED)}")
print(f"[mbench-ann] pairwise submitted: {len(PAIRWISE_SUBMITTED)}")
print(f"[mbench-ann] MBench-A completed: {sum(len(v) for v in MBENCH_A_COMPLETED.values())} annotations across {len(MBENCH_A_COMPLETED)} tasks")
# ---------------------------------------------------------------------------
# Queue helpers
# ---------------------------------------------------------------------------
def _reap_expired(pending_dict):
now = time.time()
expired = [k for k, (_, ts) in pending_dict.items() if now - ts > PENDING_TIMEOUT_SEC]
for k in expired:
pending_dict.pop(k, None)
def _append_annotation(record: dict[str, Any], ann_file: Path) -> None:
line = json.dumps(record, ensure_ascii=False)
if scheduler is not None:
with scheduler.lock:
with ann_file.open("a", encoding="utf-8") as f:
f.write(line + "\n")
else:
with ann_file.open("a", encoding="utf-8") as f:
f.write(line + "\n")
# ---------------------------------------------------------------------------
# Binary annotation callbacks (MBench-V)
# ---------------------------------------------------------------------------
def binary_start(annotator: str, state: dict):
annotator = (annotator or "").strip()
if not annotator:
return state, "<p>请先输入名字。</p>", "", "", "⚠️ 请输入名字", ""
order = list(range(len(BINARY_POOL)))
random.shuffle(order)
state = {"annotator": annotator, "order": order, "idx": 0, "current": None, "count": 0}
return _binary_next(state)
def _binary_next(state):
annotator = state["annotator"]
order = state["order"]
idx = state.get("idx", 0)
with STATE_LOCK:
_reap_expired(BINARY_PENDING)
while idx < len(order):
mt = BINARY_POOL[order[idx]]
if mt in BINARY_SUBMITTED or mt in BINARY_PENDING:
idx += 1
continue
BINARY_PENDING[mt] = (annotator, time.time())
state["idx"] = idx
state["current"] = mt
model, task_id = mt
task = TASK_BY_ID[task_id]
video_html = _render_video_html(_video_url(model, task_id))
meta = f"**模型**: `{model}` | **task_id**: `{task_id}` | **已提交**: {state['count']}"
prompt = _extract_prompt(task)
n_sub = len(BINARY_SUBMITTED)
stats = f"全局进度: {n_sub}/{len(BINARY_POOL)} ({100*n_sub/len(BINARY_POOL):.1f}%)"
return state, video_html, meta, prompt, f"✅ 已加载", stats
state["current"] = None
return state, "<p>🎉 全部完成!</p>", "全部标注完成", "", "完成", f"已完成 {len(BINARY_SUBMITTED)}/{len(BINARY_POOL)}"
def binary_submit(state, verdict, note):
if not state or not state.get("current"):
return state, "<p>请先登录</p>", "", "", "否", "", "⚠️", ""
mt = state["current"]
model, task_id = mt
record = {
"type": "binary",
"timestamp": time.time(),
"annotator": state["annotator"],
"model": model,
"task_id": task_id,
"memory_issue": verdict == "是",
"verdict": verdict,
"note": (note or "").strip(),
}
_append_annotation(record, ANN_FILE_BINARY)
with STATE_LOCK:
BINARY_PENDING.pop(mt, None)
BINARY_SUBMITTED.add(mt)
state["count"] = state.get("count", 0) + 1
state["idx"] = state["idx"] + 1
state["current"] = None
result = _binary_next(state)
return result[0], result[1], result[2], result[3], "否", "", f"✅ 已提交第 {state['count']} 条", result[5]
def binary_skip(state):
if not state or not state.get("current"):
return state, "<p>请先登录</p>", "", "", "否", "", "⚠️", ""
mt = state["current"]
with STATE_LOCK:
BINARY_PENDING.pop(mt, None)
state["idx"] = state["idx"] + 1
state["current"] = None
result = _binary_next(state)
return result[0], result[1], result[2], result[3], "否", "", "⏭️ 已跳过", result[5]
# ---------------------------------------------------------------------------
# MBench-V Pairwise annotation callbacks
# ---------------------------------------------------------------------------
def pairwise_start(annotator: str, dimension: str, state: dict):
annotator = (annotator or "").strip()
if not annotator:
return state, "<p>请先输入名字。</p>", "<p></p>", "", "", "⚠️ 请输入名字", ""
dim_pool = [(i, item) for i, item in enumerate(PAIRWISE_POOL) if item[3] == dimension]
order = list(range(len(dim_pool)))
random.shuffle(order)
state = {
"annotator": annotator, "dimension": dimension, "dim_pool": dim_pool,
"order": order, "idx": 0, "current": None, "count": 0,
}
return _pairwise_next(state)
def _pairwise_next(state):
annotator = state["annotator"]
dim_pool = state["dim_pool"]
order = state["order"]
idx = state.get("idx", 0)
dimension = state["dimension"]
dim_label = dimension
dim_question = ""
for dk, dl, dq in PAIRWISE_DIMENSIONS:
if dk == dimension:
dim_label = dl
dim_question = dq
break
with STATE_LOCK:
_reap_expired(PAIRWISE_PENDING)
while idx < len(order):
pool_idx, item = dim_pool[order[idx]]
tid, m_a, m_b = item[0], item[1], item[2]
if item in PAIRWISE_SUBMITTED or item in PAIRWISE_PENDING:
idx += 1
continue
PAIRWISE_PENDING[item] = (annotator, time.time())
state["idx"] = idx
state["current"] = item
if random.random() < 0.5:
left_model, right_model = m_a, m_b
state["swapped"] = False
else:
left_model, right_model = m_b, m_a
state["swapped"] = True
task = TASK_BY_ID[tid]
video_a_html = _render_video_html(_video_url(left_model, tid))
video_b_html = _render_video_html(_video_url(right_model, tid))
prompt = _extract_prompt(task)
meta = f"**维度**: {dim_label} | **问题**: {dim_question}\n\n**已提交**: {state['count']}"
n_sub = sum(1 for x in PAIRWISE_SUBMITTED if x[3] == dimension)
n_total = len(dim_pool)
stats = f"维度「{dim_label}」进度: {n_sub}/{n_total} ({100*n_sub/n_total:.1f}%)"
return state, video_a_html, video_b_html, meta, prompt, "✅ 已加载", stats
state["current"] = None
return state, "<p>🎉 该维度全部完成!</p>", "", "全部完成", "", "完成", ""
def pairwise_submit(state, verdict, note):
if not state or not state.get("current"):
return state, "", "", "", "", "⚠️ 请先登录", ""
item = state["current"]
tid, m_a, m_b, dimension = item
swapped = state.get("swapped", False)
if verdict == "左边更好":
winner = m_b if swapped else m_a
elif verdict == "右边更好":
winner = m_a if swapped else m_b
else:
winner = "tie"
record = {
"type": "pairwise",
"timestamp": time.time(),
"annotator": state["annotator"],
"task_id": tid,
"model_a": m_a,
"model_b": m_b,
"dimension": dimension,
"winner": winner,
"verdict_raw": verdict,
"swapped": swapped,
"note": (note or "").strip(),
}
_append_annotation(record, ANN_FILE_PAIRWISE)
with STATE_LOCK:
PAIRWISE_PENDING.pop(item, None)
PAIRWISE_SUBMITTED.add(item)
state["count"] = state.get("count", 0) + 1
state["idx"] = state["idx"] + 1
state["current"] = None
result = _pairwise_next(state)
return result[0], result[1], result[2], result[3], result[4], f"✅ 已提交第 {state['count']} 条", result[6]
def pairwise_skip(state):
if not state or not state.get("current"):
return state, "", "", "", "", "⚠️ 请先登录", ""
item = state["current"]
with STATE_LOCK:
PAIRWISE_PENDING.pop(item, None)
state["idx"] = state["idx"] + 1
state["current"] = None
result = _pairwise_next(state)
return result[0], result[1], result[2], result[3], result[4], "⏭️ 已跳过", result[6]
# ---------------------------------------------------------------------------
# MBench-A Pairwise annotation callbacks
# ---------------------------------------------------------------------------
def mbench_a_start(annotator: str, state: dict):
"""Login for MBench-A annotation."""
annotator = (annotator or "").strip()
if not annotator:
return (state, "⚠️ 请输入名字", "", "", "", "",
gr.update(visible=False), gr.update(visible=False),
gr.update(visible=False), gr.update(visible=False),
gr.update(visible=False),
"", "")
# Count how many tasks this annotator has already completed.
# Check both:
# 1. MBENCH_A_COMPLETED (loaded from HF at startup + updated in-memory during this session)
# 2. The local annotation file (captures annotations made this session before any push)
historical_count = sum(
1 for anns in MBENCH_A_COMPLETED.values()
if annotator in anns
)
# Also scan the local file in case this session's annotations haven't been pushed yet
if ANN_FILE_MBENCH_A.exists():
with ANN_FILE_MBENCH_A.open() as f:
for line in f:
line = line.strip()
if not line:
continue
try:
r = json.loads(line)
if r.get("annotator") == annotator and r.get("type") == "pairwise_mbench_a":
tid = r.get("task_id", "")
# Only count if not already counted in MBENCH_A_COMPLETED
if tid in MBENCH_A_TASK_BY_ID and annotator not in MBENCH_A_COMPLETED.get(tid, []):
historical_count += 1
except Exception:
pass
# Shuffle task order for this annotator
order = list(range(len(MBENCH_A_TASKS)))
random.shuffle(order)
state = {
"annotator": annotator,
"order": order,
"idx": 0,
"current_task_id": None,
"swapped": False,
"left_model": None,
"right_model": None,
"count": historical_count,
}
return _mbench_a_next(state)
def _mbench_a_next(state: dict):
"""Find and load the next available MBench-A task."""
annotator = state["annotator"]
order = state["order"]
idx = state.get("idx", 0)
with STATE_LOCK:
_reap_expired(MBENCH_A_PENDING)
while idx < len(order):
task = MBENCH_A_TASKS[order[idx]]
tid = task["task_id"]
# Skip if already fully annotated
if len(MBENCH_A_COMPLETED.get(tid, [])) >= MBENCH_A_ANNOTATORS_PER_TASK:
idx += 1
continue
# Skip if this annotator already did it
if annotator in MBENCH_A_COMPLETED.get(tid, []):
idx += 1
continue
# Skip if currently pending by someone else
if tid in MBENCH_A_PENDING and MBENCH_A_PENDING[tid][0] != annotator:
idx += 1
continue
# Assign this task
MBENCH_A_PENDING[tid] = (annotator, time.time())
state["idx"] = idx
state["current_task_id"] = tid
# Randomly swap A/B
m_a, m_b = task["model_a"], task["model_b"]
if random.random() < 0.5:
state["left_model"], state["right_model"] = m_a, m_b
state["swapped"] = False
else:
state["left_model"], state["right_model"] = m_b, m_a
state["swapped"] = True
# Build UI outputs
subset = task["subset"]
video_left = _render_video_html(
_mbench_a_video_proxy_url(state["left_model"], subset, task["sample_id"]))
video_right = _render_video_html(
_mbench_a_video_proxy_url(state["right_model"], subset, task["sample_id"]))
aux_html = _render_mbench_a_aux(task)
# Dimension questions
dimensions = task["dimensions"]
dim_questions = task.get("dimension_questions", {})
# Build question radio updates (max 5)
q_updates = []
for i in range(6):
if i < len(dimensions):
dim_key = dimensions[i]
question_text = dim_questions.get(dim_key, dim_key)
q_updates.append(gr.update(
visible=True,
label=question_text,
value="差不多",
))
else:
q_updates.append(gr.update(visible=False, value="差不多"))
# Meta info
subset_names = {"environment": "🏞️ Environment", "object": "🎯 Object",
"human": "👤 Human", "causal": "⚡ Causal"}
n_done = sum(1 for t in MBENCH_A_TASKS
if len(MBENCH_A_COMPLETED.get(t["task_id"], [])) >= MBENCH_A_ANNOTATORS_PER_TASK)
meta = (f"**子集**: {subset_names.get(subset, subset)} | "
f"**已提交**: {state['count']}")
stats = (f"全局进度: {n_done}/{len(MBENCH_A_TASKS)} tasks 完成 | "
f"你已标注: {state['count']}")
return (state, "✅ 已加载", aux_html, video_left, video_right, meta,
*q_updates, "", stats)
# All done
state["current_task_id"] = None
empty_q = gr.update(visible=False, value="差不多")
return (state, "🎉 全部完成!", "", "<p>所有任务已完成</p>", "", "全部完成",
empty_q, empty_q, empty_q, empty_q, empty_q, empty_q, "", "")
def mbench_a_submit(state, q1_val, q2_val, q3_val, q4_val, q5_val, q6_val, note):
"""Submit MBench-A multi-dimension annotation."""
if not state or not state.get("current_task_id"):
empty_q = gr.update(visible=False, value="差不多")
return (state, "⚠️ 请先登录", "", "", "", "",
empty_q, empty_q, empty_q, empty_q, empty_q, empty_q, "", "")
tid = state["current_task_id"]
task = MBENCH_A_TASK_BY_ID[tid]
dimensions = task["dimensions"]
swapped = state["swapped"]
m_a, m_b = task["model_a"], task["model_b"]
# Map verdicts to winners
verdicts = [q1_val, q2_val, q3_val, q4_val, q5_val, q6_val]
dim_results = {}
for i, dim_key in enumerate(dimensions):
v = verdicts[i]
if v == "A更好":
# A is left; if swapped, left is model_b
winner = m_b if swapped else m_a
elif v == "B更好":
winner = m_a if swapped else m_b
else:
winner = "tie"
dim_results[dim_key] = winner
record = {
"type": "pairwise_mbench_a",
"timestamp": time.time(),
"annotator": state["annotator"],
"task_id": tid,
"subset": task["subset"],
"sample_id": task["sample_id"],
"camera_motion": task.get("camera_motion", "left_then_right"),
"model_a": m_a,
"model_b": m_b,
"dimensions": dim_results,
"swapped": swapped,
"note": (note or "").strip(),
}
_append_annotation(record, ANN_FILE_MBENCH_A)
with STATE_LOCK:
MBENCH_A_PENDING.pop(tid, None)
MBENCH_A_COMPLETED[tid].append(state["annotator"])
state["count"] = state.get("count", 0) + 1
state["idx"] = state["idx"] + 1
state["current_task_id"] = None
return _mbench_a_next(state)
def mbench_a_skip(state):
"""Skip current MBench-A task."""
if not state or not state.get("current_task_id"):
empty_q = gr.update(visible=False, value="差不多")
return (state, "⚠️ 请先登录", "", "", "", "",
empty_q, empty_q, empty_q, empty_q, empty_q, empty_q, "", "")
tid = state["current_task_id"]
with STATE_LOCK:
MBENCH_A_PENDING.pop(tid, None)
state["idx"] = state["idx"] + 1
state["current_task_id"] = None
return _mbench_a_next(state)
# ---------------------------------------------------------------------------
# UI
# ---------------------------------------------------------------------------
CUSTOM_CSS = """
#prompt_box textarea { height: 300px !important; overflow-y: auto !important; }
.video-pair { display: flex; gap: 12px; }
.video-pair > div { flex: 1; }
/* Force aux info box to be visible regardless of Gradio theme */
.aux-info-box {
background: #e3e8ef !important;
color: #111 !important;
padding: 14px !important;
border-radius: 8px !important;
margin-bottom: 12px !important;
border: 1px solid #b0b8c4 !important;
}
.aux-info-box * {
color: #111 !important;
}
.aux-info-box img {
border: 1px solid #999;
border-radius: 4px;
}
"""
with gr.Blocks(title="MBench 标注", theme=gr.themes.Soft(), css=CUSTOM_CSS) as demo:
gr.Markdown("# 🎬 MBench 视频标注平台")
with gr.Tabs():
# ═══════════════ MBench-A Pairwise ═══════════════
with gr.Tab("MBench-A 对比 (World Models)"):
gr.Markdown(
"## 🌍 MBench-A — 世界模型记忆能力评测\n\n"
"比较两个世界模型生成的长视频(~25 秒),评估相机转走再转回来后的记忆一致性。\n\n"
"**视频 A/B 的模型身份已匿名随机分配。请对每个维度独立判断。**"
)
a_stats = gr.Markdown("")
a_state = gr.State({})
with gr.Row():
a_name = gr.Textbox(label="标注员名字", placeholder="例如: charlie", scale=4)
a_login = gr.Button("开始标注", variant="primary", scale=1)
a_status = gr.Markdown("")
# Auxiliary info (mask image / camera GIF + caption / instructions)
a_aux = gr.HTML("")
# Video pair
with gr.Row(equal_height=True):
with gr.Column(scale=1, min_width=360):
gr.Markdown("### 视频 A")
a_video_left = gr.HTML("<p>请先登录。</p>")
with gr.Column(scale=1, min_width=360):
gr.Markdown("### 视频 B")
a_video_right = gr.HTML("<p>请先登录。</p>")
# Task info
a_meta = gr.Markdown("")
# Multi-dimension questions (max 6, dynamically shown/hidden)
gr.Markdown("---\n### 请对以下每个维度分别判断:")
a_q1 = gr.Radio(["A更好", "差不多", "B更好"], value="差不多", label="维度 1", visible=False)
a_q2 = gr.Radio(["A更好", "差不多", "B更好"], value="差不多", label="维度 2", visible=False)
a_q3 = gr.Radio(["A更好", "差不多", "B更好"], value="差不多", label="维度 3", visible=False)
a_q4 = gr.Radio(["A更好", "差不多", "B更好"], value="差不多", label="维度 4", visible=False)
a_q5 = gr.Radio(["A更好", "差不多", "B更好"], value="差不多", label="维度 5", visible=False)
a_q6 = gr.Radio(["A更好", "差不多", "B更好"], value="差不多", label="维度 6", visible=False)
a_note = gr.Textbox(label="备注(可选)", lines=1)
with gr.Row():
a_submit = gr.Button("✅ 提交并下一组", variant="primary")
a_skip = gr.Button("⏭️ 跳过")
# Wiring
a_all_outs = [a_state, a_status, a_aux, a_video_left, a_video_right, a_meta,
a_q1, a_q2, a_q3, a_q4, a_q5, a_q6, a_note, a_stats]
a_login.click(mbench_a_start, [a_name, a_state], a_all_outs)
a_name.submit(mbench_a_start, [a_name, a_state], a_all_outs)
a_submit.click(mbench_a_submit,
[a_state, a_q1, a_q2, a_q3, a_q4, a_q5, a_q6, a_note], a_all_outs)
a_skip.click(mbench_a_skip, [a_state], a_all_outs)
# ---------------------------------------------------------------------------
# Video proxy
# ---------------------------------------------------------------------------
if __name__ == "__main__":
import httpx
from fastapi import HTTPException, Request
from fastapi.responses import StreamingResponse
from gradio.routes import App as _GradioApp
_video_client = httpx.AsyncClient(timeout=30.0, follow_redirects=True)
async def _do_proxy(upstream: str, request: Request):
"""Generic proxy for HF video/asset URLs."""
req_headers = {}
if (rng := request.headers.get("range")):
req_headers["range"] = rng
try:
upstream_resp = await _video_client.send(
_video_client.build_request("GET", upstream, headers=req_headers),
stream=True,
)
except Exception as e:
raise HTTPException(502, f"upstream fetch failed: {e}")
passthrough_headers = {}
for h in ("content-type", "content-length", "accept-ranges",
"content-range", "etag", "last-modified"):
if h in upstream_resp.headers:
passthrough_headers[h] = upstream_resp.headers[h]
passthrough_headers.setdefault("content-type", "video/mp4")
passthrough_headers["cache-control"] = "public, max-age=300"
async def _body():
try:
async for chunk in upstream_resp.aiter_bytes(chunk_size=65536):
yield chunk
finally:
await upstream_resp.aclose()
return StreamingResponse(_body(), status_code=upstream_resp.status_code, headers=passthrough_headers)
async def _proxy_video(model: str, task_id: str, request: Request):
"""Proxy MBench-V videos."""
if model not in MODELS or task_id not in TASK_BY_ID:
raise HTTPException(404, "unknown (model, task_id)")
upstream = _hf_video_url(model, task_id)
return await _do_proxy(upstream, request)
async def _proxy_mbench_a_video(model: str, category: str, sample_id: str, request: Request):
"""Proxy MBench-A videos."""
if model not in MBENCH_A_MODELS:
raise HTTPException(404, f"unknown model: {model}")
upstream = _mbench_a_hf_video_url(model, category, sample_id)
return await _do_proxy(upstream, request)
_orig_create_app = _GradioApp.create_app
def _patched_create_app(*args, **kwargs):
app = _orig_create_app(*args, **kwargs)
# MBench-V video proxy
app.add_api_route(
"/video/{model}/{task_id}.mp4",
_proxy_video,
methods=["GET", "HEAD"],
include_in_schema=False,
)
# MBench-A video proxy
app.add_api_route(
"/video_a/{model}/{category}/{sample_id}/left_then_right.mp4",
_proxy_mbench_a_video,
methods=["GET", "HEAD"],
include_in_schema=False,
)
print("[mbench-ann] video proxy routes registered (MBench-V + MBench-A)")
return app
_GradioApp.create_app = staticmethod(_patched_create_app)
demo.queue(default_concurrency_limit=16).launch(ssr_mode=False)