# ASR-Bench-2k v3 ## ASR-Bench-1k — 2026-09-08 用户确认将审阅后的 benchmark 调整为1000条，接受比例变化，不再为旧配额补题。 这是原 ASR-Bench-2k 的隔离子集；原2k数据、审阅记录均保留，未覆盖。 ### 产物与使用 根目录：`/scratch/projects/CFP03/CFP03-CF-035/zhaoyanguang/data/ASR-Bench-1k` - `release/benchmark_public_v3.jsonl`：1000条模型输入；媒体路径相对根目录。 - `release/benchmark_hidden_v3.jsonl`：1000条答案、绑定及金标信息，仅供评测器使用。 - `release/coordinate_train_v3.jsonl`：原格式完整参考记录；虽沿用文件名，本数据是val benchmark，不用于训练。 - `images/`、`videos/`：902个实际复制的去重媒体文件，无原目录链接依赖。 - `sidecar_v3/`：样本相关表过滤为1000条；视频索引仅保留引用视频。场景几何和逐帧对象库保留父版支持数据超集，便于核验，不计入题量。 - `provenance/`：完整2k审阅、选中1000条、排除4条及原因、结构样本、v5中间记录、父版验收报告。 - `manifest.json`：本子集验收与分布。父版验收报告明确标为2k历史报告，不伪装成新1k验收。 ### 选择与最终分布 只使用 `usable` / `usable_note` 的1004条；去掉审阅编号51、90、315、1272。 这4条均为单场景图片尺度题，同款搬家箱的单数指代或左右排序存在轻微问题，内容重复度较高。 不将其重新判为重大错误；完整原审阅结论保留。全部跨场景、视频、距离题保留。 去除前后场景集合相同，仍为50个场景。 | 维度 | 数量 | 比例 | |---|---:|---:| | 跨场景给定参照 | 573 | 57.3% | | 单场景给定参照 | 209 | 20.9% | | 单场景自选参照 | 218 | 21.8% | | 图片 | 751 | 75.1% | | 视频 | 249 | 24.9% | | 距离 | 744 | 74.4% | | 尺度 | 256 | 25.6% | 距离D1/D2/D3为346/347/51；尺度S1/S2/S3为102/52/102，所有任务类别保留。 这是接受后的实际分布，不再要求旧60/40、70/30、80/20配额。 ### 验证边界 发布记录按sample_id过滤；public/hidden/reference和样本sidecar一一对应，均1000唯一ID。 QA、答案、reasoning v3、对象绑定及16帧采样内容不改写。媒体复制核对大小，沿用父版解码验收，不声称重跑全量解码。 视觉质量依据是答案可见的模型逐条审阅，不是独立人工标注，也不是重新测量GT；可用不代表保证数值真值无误。 未启动训练或push。Project_Storage复制状态见下文。构建脚本：`runtime/publish_asr_bench1.py`，拒绝覆盖已有目标。 ### Project_Storage 复制目标：`/Project_Storage/CFP-03/CFP03-CF-035/zhaoyanguang/data/AnchorSR/ASR-Bench-1k`。 完整复制发布目录（媒体、QA、Sidecar和溯源记录），本机数据保留。当前状态：复制完成。 2026-09-08：1075个文件，约584 MiB；rsync成功结束，随后核对文件路径/大小及1000条样本对应关系。 以下为原2k构建与审阅历史。 ## Latest visual QA qualification — 2026-09-08 Full 2,000-row visual QA review is **complete** (1,400 image / 600 video), using the same rubric below. The original 100 judgments were carried over after exact QA/media comparison by sample ID; the other 1,900 were visually inspected. Results: **1,004 usable or minor-note (50.20%), 451 major (22.55%), 545 unresolved (27.25%)**. Unresolved does not mean proven wrong; this is answer-visible model review, not independent human labeling or GT remeasurement. Coverage checks passed with no missing or duplicate samples. No release data was modified. This is review completion, **not unconditional release approval**. See [review protocol and checkpoint](../../runtime/bench_qa_review2000/REVIEW.md) and [live counts](../../runtime/bench_qa_review2000/progress.json). The final release's automatic acceptance is **not** a final visual QA approval. At the user's request, Codex visually inspected a fixed stratified 100-sample draw: 70 image / 30 video, 60 cross-scene / 40 single-scene, 80 distance / 20 scale, covering 43 scenes. Each video input was reviewed at all 16 declared uniform16 frames. Only input media, question and answer were evaluated; gold reasoning was excluded. Minor wording, partial visibility and normal estimation difficulty were not automatic grounds for rejection. Results: **56 usable/no major issue observed** (28 with minor notes), **25 material QA defects** (17 ambiguous targets, 8 unstable video ordinal references), and **19 unresolved cases**. Images: 46 / 12 / 12; videos: 10 / 13 / 7, respectively. These are model-review judgments, not independent human labels or remeasured metric-GT accuracy. The review was answer-visible, not blinded; unresolved cases were counted neither as passes nor proven errors. The material defects primarily concern question interpretation, not proof that the stored numerical answers were computed incorrectly: e.g. multiple chairs but only “chair”, or a moving video with “eighth-from-left chair” and no reference frame. The 100-row review therefore does **not** support unconditional final benchmark-quality sign-off. No release data, media, Sidecar, reasoning, or automatic acceptance manifest was changed. Question repairs / replacements require a subsequent authorized task; neither Train nor image-base was rerun. Full per-sample judgments, IDs, questions, answers, media paths and visual evidence: [100-row QA review](../../runtime/bench_qa_review100/REVIEW.md). Machine-readable records: [review_results.json](../../runtime/bench_qa_review100/review_results.json). The preparation script is `runtime/bench_review100.py`; sampling seed 20260909. The summarizer only packages manually authored judgments, not automated scores. ## Goal and isolation ASR-Bench-2k is a quality-first 2,000-sample benchmark built exclusively from the 50 official ScanNet++ `nvs_sem_val` scenes. It is isolated from the 160k Train release and from all earlier Bench rehearsals. - raw ScanNet++ root: `/scratch/projects/CFP03/CFP03-CF-035/zhaoyanguang/data/scannet` - formal v3 root: `/scratch/projects/CFP03/CFP03-CF-035/zhaoyanguang/data/asr_bench_2k_v3_coordinate_uniform16_v1` - current smoke root: `/scratch/projects/CFP03/CFP03-CF-035/zhaoyanguang/data/asr_bench_2k_v3_coordinate_uniform16_v1_smoke6` - old `asr_bench_2k_v1*` roots: historical regression fixtures only; never a parent of the formal v3 release ## Frozen target | Axis | Exact count | |---|---:| | image / video | 1,400 / 600 | | single-scene / cross-scene | 800 / 1,200 | | given reference / self-selected | 1,600 / 400 | | distance / scale | 1,600 / 400 | All 50 Val scenes must occur at least 20 times, and cross-scene samples must cover at least 120 distinct scene pairs. Selection remains quality-first and deterministic with fact, anchor, scene, scene-pair, answer-bin, and video-asset reuse bounds. Cross-scene T2/T3 graphs are outside the supported task contract. ## Construction order ```text official Val raw assets → image-base scene shards → image composition + video atomic (parallel) → video composition → taxonomy supply audit → exact global 2k selection → media materialization and legacy structural validation → Sidecar-v1 semantic graph → full-frame video trajectories + 3D OBB + typed distance metrics → exact uniform16 video sampling → coordinate grounding and frozen object bindings → v2.5 natural question/object wording → canonical reasoning v3 prompt and symbolic graph → v3 blocking acceptance ``` Legacy question/reasoning text is only structural scaffolding. It is parsed against the exact parent Sidecar that licensed it, while repaired labels, uniform16 phrases, coordinates, and final v3 validation use the new Sidecar. This prevents old wording from leaking into publication without weakening the new coordinate-bound parser. ## Published contract - public records contain only media metadata and the system/user prompt; taxonomy, answers, object IDs, Sidecar Gold, gold reasoning, and grounding counts are not published to the model-facing file; - hidden records contain answers, v3 reasoning, authoritative object bindings, exact uniform16 mappings, Sidecar-v3 trajectories/geometry, and provenance; - each symbolic `Object` must match exactly one visible Sidecar object with IoU at least 0.99 and a positive uniqueness margin; - every video uses `uniform16-linspace-floor-v1`: `floor(i*(N-1)/15), i=0..15`; - v3 Parser/Sidecar/Reward results must be identical to the frozen semantic parent. Natural measurement quantization is accepted; semantic drift is not; - all media must exist and decode, all inputs must be Val, quotas and coverage must be exact, and duplicate model-visible inputs are forbidden. ## Resumability and observability Core DAG state is written under `runtime/construction_progress.json`, `runtime/full_construction_manifest.json`, and `runtime/stage_progress/`. Sidecar-v1 and v3 generation use SQLite checkpoints plus atomic progress JSON. Checkpoint reuse is bound to source hashes, release identity, video sampling, and an implementation signature; changed code cannot silently reuse stale records. The v3 bundle writes: - `v3/release/benchmark_public_v3.jsonl` - `v3/release/benchmark_hidden_v3.jsonl` - `v3/release/coordinate_train_v3.jsonl` (gold reference, not public input) - `v3/sidecar_v3/` - `v3/audit/` with dedup, coverage, media, schema, Sidecar, uniform16, leakage, and final acceptance reports - `v3/accepted_manifest.json` ## Execution gate and current status The six-scene smoke has the exact scaled 60-row joint distribution: 42 image, 18 video, 36 cross-scene, 24 single-scene, 48 distance, and 12 scale. A downstream-only 60-row regression passed 60/60 and all 240 Object declarations; the complete `data_pipeline/tests` suite passes 303/303. Two compatibility defects found before formal production are now covered by regressions: legacy wording is parsed against its parent Sidecar after uniform16 resampling, and the ScanNet++ `carboard box` spelling variant is canonicalized consistently in video relational grounding. Both fail closed rather than producing ambiguous v3 data. The 60-row v3 smoke bundle passed automatic and manual review. Formal Val candidate manufacture then completed for all 50 official scenes. The first global selection and a protected-reservoir replay both stopped safely before materialisation because the frozen 2,000-row contract was not met. The latest incomplete result is frozen as **incomplete v0** at: `runtime/checkpoints/incomplete_v0_20260907/` It contains the exact selector source snapshot, 30,353 retained candidates, the 1,631 selected rows, progress state, and the full selection report. It is recoverable diagnostic state, not a publishable benchmark. Its remaining shortfall is 369 cross-scene/given-reference/distance rows: 231 image and 138 video. Forty-nine scenes are represented; `13c3e046d7` has 17/20 incidences and `a980334473` has 0/20. The latter scene's raw and image-base stages are complete. A targeted audit found that its original zero supply came from the per-fact top-8 frame preselection: eligible anchors first appeared later in the ranked frame list. Expanding only this search to 256 frames, without relaxing any geometric or quality threshold, recovered 17 routed candidates covering 14 distinct facts. The protected replay retained every structurally valid, non-denylisted cross-scene distance candidate (4,522 image and 2,092 video), yet improved the greedy selection only from 1,596 to 1,631. Reservoir truncation was therefore not the remaining blocker. An exact MILP audit proved that candidate supply and the 0.40 quality floor were sufficient, while the Train-scale combination of `max_samples_per_scene_pair=10` and `max_uses_per_target_fact=2` was jointly infeasible for a 60%-cross-scene benchmark over only 50 Val scenes. The fact bound counts both source facts of each cross-scene composition; it is not just a bound on the final answer fact. The recovery lineage uses a Bench-only `quality_score >= 0.40` hard floor and then maximises quality subject to the complete contract. This is materially stricter than the accepted Train-160k distribution: Train has mean 0.4596, median 0.4545 and 66,232/160,000 rows below 0.40. Incomplete v0 already has mean 0.7027 and median 0.7936, but contains 256 rows below 0.40; those rows are not eligible for the recovered release. The joint selection report records overall and per-cell quality quantiles so the final comparison does not depend on a different task mixture. A pre-solver necessary-condition audit leaves over 25k candidates after the 0.40 floor. Every one of the 12 exact task cells remains oversupplied; the tightest is image/cross-scene/given-reference/scale at 410 candidates for 168 slots (2.44x). The two formerly short cross-scene distance cells retain 3,502 image candidates for 672 slots and 1,536 video candidates for 288 slots. All 49 already represented scenes have at least 61 quality-valid incidences; the 50th scene is supplied by the isolated frame-search recovery. Aggregate answer-bin, scene, scene-pair, fact, anchor and video-asset capacity upper bounds all exceed the corresponding release demand. These were necessary, not sufficient, conditions; the joint solver was the authoritative test of their overlap. Recovery is isolated from v0. The recovered scene candidates are composed in their own runtime directory and combined with the immutable v0 reservoir by a MILP selector that jointly enforces all 12 exact cells, all 50 scene minima, deduplication, fact-evidence separation and every diversity bound. A systematic boundary search established the minimum feasible Bench-only pair as `max_samples_per_scene_pair=11` and `max_uses_per_target_fact=11`; either bound at 10 remains infeasible even when the other is made non-binding. No geometric, structural, deduplication, evidence-separation, or quality threshold was relaxed, and the Train configuration is unchanged. The formal joint result at `base/joint_selection_v1` is accepted for materialisation: 2,000/2,000 rows, exact 12-cell quotas, all 50 scenes with at least 20 incidences, 273 distinct cross-scene pairs, and zero validation violations. Quality is minimum 0.4004, mean 0.7357, and median 0.7769. The final selector also treats the canonical model-visible media identity plus question as a hard uniqueness key. This closes a cross-branch collision in which atomic/routed and composition schemas could encode the same eventual model input differently. The corrected selection retains 1,861 rows from the first joint solution, replaces 139, and has zero duplicate model-visible inputs. The failed greedy `base/global_selection`, incomplete v0 checkpoint, and pre-correction joint result remain diagnostic archives and are never used as publishable outputs. Formal entrypoints remain: - `data_pipeline/jobs/scannetpp_asr_bench_2k_v3_full_core.pbs` - `data_pipeline/jobs/scannetpp_asr_bench_2k_v3_full_finalize.pbs` - `data_pipeline/jobs/scannetpp_asr_bench_v3_a980_compose_and_solve.pbs` - `data_pipeline/jobs/scannetpp_asr_bench_v3_joint_materialize.pbs` - `data_pipeline/jobs/scannetpp_asr_bench_v3_joint_finalize.pbs` The formal config is `ready_for_full_benchmark`. Construction and all blocking acceptance gates completed on 2026-09-08. The accepted release root is `/scratch/projects/CFP03/CFP03-CF-035/zhaoyanguang/data/asr_bench_2k_v3_coordinate_uniform16_v1/v3`; `accepted_manifest.json` is the authoritative entrypoint. It records 2,000 records (1,400 image and 600 video), 1,200 cross-scene and 800 single-scene questions, 1,600 distance and 400 scale questions, 8,025 object declarations, and 1,648 decoded unique media assets. All nine audit reports pass: model-input deduplication, leakage, media decoding/path safety, schema/parser-reward equivalence, Sidecar v3, uniform16 sampling, scene coverage, cross-scene graph connectivity, and aggregate v3 acceptance. In particular, duplicate model-visible inputs, duplicate sample IDs, gold leakage, missing media, undecodable media, and uniform16 violations are all zero.