{"id":"0182db65-4382-4ce2-80d6-2d73c21a7247","arxiv_id":"2504.15182","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Tiger200K is a manually curated dataset of 170,000 bilingual-captioned video clips from BiliBili, claimed to have high visual quality for fine-tuning video generation models.","lead":"A researcher collected and manually curated 170,000 short video clips from a Chinese video platform, aiming to give AI video generators high-quality training data. The paper describes the filtering pipeline and captions, but does not yet release the data or show that the videos improve generation quality.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim that Tiger200K has superior visual quality is supported only by the author's own manual curation and random self-review; no external quality metric or inter-annotator agreement is reported, so the advantage over Koala-36M is unverified.","rationale":"I read the paper as a dataset contribution whose value depends on two things: the data being released and the data being measurably higher quality than existing alternatives. The pipeline is described in sufficient detail and the fixed-length cut design (Section 2.2) plausibly supports temporal consistency, which is a genuine strength. However, the quality advantage is asserted rather than demonstrated. The reader's weakest assumption identifies exactly this gap, and I agree. I do not see an internal inconsistency that would warrant rejection; the missing evidence is an empirical validation task, not a logical flaw. The CONDITIONAL verdict is therefore appropriate: acceptance should require dataset release and an objective quality evaluation. My stress-test does not change that verdict.","tokens_in":6965,"tokens_out":3151,"duration_ms":28127,"concrete_test":"Release the dataset with a persistent identifier and run a double-blind human-preference comparison: randomly sample 500 clips from Tiger200K and 500 from Koala-36M (matched in duration and resolution), have at least 10 independent raters score visual quality, aesthetics, and temporal consistency without knowing the source, and report mean scores plus inter-annotator agreement (e.g., Krippendorff's alpha). If Tiger200K does not significantly outperform Koala-36M on these dimensions, the central quality claim is not supported. As a complementary check, fine-tune a fixed video-generation baseline on equal-size Tiger200K and Koala-36M subsets and compare outputs; the absence of a downstream win would further weaken the claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Tiger200K provides high visual quality, temporally consistent video-text pairs suitable for post-training (Abstract, Section 3, Conclusion). The load-bearing assumption is that the author's manual curation and manual random sampling (Sections 2.1 and 2.4) are sufficient evidence of quality. This assumption is not tested anywhere in the paper. No quantitative quality metric is defined, no inter-annotator agreement is reported, and no external human-preference study or downstream generation experiment is run. Section 1 asserts that Koala-36M 'fails to meet the stringent quality requirements' for post-training, and Section 4 claims 'strong competitiveness in visual quality', but the only comparative evidence offered is the author's own selected examples in Figure 1. Because the quality judgment is both subjective and self-referential, the headline claim is currently unfalsifiable. Additionally, the dataset is not released, so independent verification is impossible. This does not invalidate the pipeline description, but it means the paper's central contribution cannot yet be assessed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Tiger200K, a manually curated video dataset sourced from the UGC platform Bilibili, intended for post-training and quality-tuning of text-to-video generation models. The construction pipeline has five stages: manual curation of creators/videos based on aesthetic criteria, TransNetV2-based shot boundary detection with fixed-length cut generation, safe-zone computation using OCR and black-border detection, motion filtering and manual quality review, and bilingual (Chinese-English) dense captioning with Qwen2.5-VL. The paper reports 4,151 source videos yielding 85k scenes and 170k clips, with statistics on resolution, caption length, and safe-zone retention area. The central claim is that the dataset exhibits high visual quality and temporal consistency, with the quality advantage attributed to human curation at the input and review stages.","tokens_in":7320,"tokens_out":3854,"duration_ms":39269,"significance":"If the quality claim can be substantiated, Tiger200K would be a useful open resource for the video-generation community, potentially filling a gap left by algorithmically filtered datasets such as Koala-36M for the fine-tuning stage. The manuscript is commendably transparent about its pipeline, provides algorithmic details for safe-zone detection, and reports aggregate statistics over a large number of clips. The use of bilingual captions and the focus on 4K/UGC sources are also practically relevant. However, the significance is currently conditional: the paper's core value proposition is high visual quality, but that proposition is not tested against any external benchmark, human preference study, or downstream generation experiment, and the dataset is not yet accessible for independent verification.","major_comments":[{"comment":"The central claim that Tiger200K has 'strong competitiveness in visual quality' is supported only by the author's own manual curation and self-review via random sampling. No inter-annotator agreement, no quantitative aesthetic metric, no comparison against Koala-36M by independent human raters, and no downstream fine-tuning experiment is reported. Because the dataset's raison d'être is quality, this absence is load-bearing. Please add at least one of the following: a human preference study comparing Tiger200K clips against Koala-36M clips, a downstream video-generation fine-tuning experiment with quantitative metrics, or a reproducible quality-rating protocol with reported agreement statistics.","section":"Sections 1, 2.1, 2.4, and 4"},{"comment":"The claim that TransNetV2 'achieves consistent segmentation performance across both synthetic test videos and real-world UGC content' is based on visual timeline comparisons only. Since temporal consistency is part of the dataset's stated value, the shot-boundary detection selection should be supported by quantitative metrics such as precision, recall, and F1 on the synthetic and real test sets, including cross-dissolve transitions. The construction of the ground-truth test set should also be described.","section":"Section 2.2 and Figure 5"},{"comment":"The descriptive statistics presented do not establish 'high quality.' In particular, the statement that a safe-zone retention area above 85% 'demonstrates the high quality of the processed data' conflates the filter's output distribution with an independent measure of visual quality. Resolution, caption length, and safe-zone area are pipeline statistics, not quality evaluations. Please separate these descriptive statistics from any quality-validity evidence, or add appropriate quantitative quality measures.","section":"Section 3"},{"comment":"The dataset is announced as 'will be released,' but no release URL, sample download, or reviewer-access mechanism is provided. For a dataset paper, the contribution cannot be independently verified without access to at least a substantial sample with metadata and captions. Please provide an anonymous review link or release the data (or a representative subset) at revision time.","section":"Abstract and Conclusion"}],"minor_comments":[{"comment":"Variables X1, Y1, X2, Y2 are used in the return statement but are never defined in the pseudocode; please define them explicitly as the final safe-zone coordinates. Also, line 31 compares a normalized area against the threshold 0.5, so the units should be stated.","section":"Section 2.3, Algorithm 1"},{"comment":"The motion filter description is underspecified: 'filter out frames below a predefined threshold' does not state whether clips are dropped based on a per-frame threshold, a fraction of low-motion frames, or a clip-level aggregate. Please clarify the exact criterion and report the threshold value.","section":"Section 2.4"},{"comment":"The phrase 'black broader' should be corrected to 'black border' in the text and pseudocode comments.","section":"Section 2.3"},{"comment":"The statement that 'the quantities in the figure are relative' makes the figure hard to interpret; please either provide actual dataset sizes or remove the misleading quantitative appearance.","section":"Figure 3"},{"comment":"Please specify the exact version of Qwen2.5-VL used and include the captioning prompt or template, as caption style substantially affects reproducibility.","section":"Section 2.5"},{"comment":"The text says scenes shorter than 121 frames are discarded, then 'center-based segmentation' produces 121-frame cuts; please state explicitly whether cuts may overlap and whether every valid scene yields floor(length/121) or another number of cuts, to make the relation between 85k scenes and 170k clips clearer.","section":"Section 2.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a short dataset-description paper. My recommendation is major revision rather than rejection because the missing quality validation is additive in nature: the pipeline description is coherent, and the paper already reports useful statistics. However, if the revision does not add external validation (human preference, downstream experiments, or quantitative shot-detection metrics) and does not provide dataset access for reviewers, I would downgrade to reject, since the paper's central claim would remain unverifiable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: Tiger200K is a real new dataset resource — 170k video clips from BiliBili, bilingual captions, half in 4K, with a clearly documented pipeline. The central claim that it is 'high visual quality' is asserted rather than demonstrated. You'll want to read this if you work on post-training data for video generation, but keep expectations calibrated until the dataset is out.\n\nWhat's genuinely new: the source is BiliBili rather than YouTube, the selection is human-curated at the creator level, and the captions are bilingual. The pipeline is described in enough detail to reproduce: TransNetV2 for shot detection, PaddleOCR plus border scanning for safe zones, Farneback optical flow for motion filtering, Qwen2.5-VL for captions. The shot-detection comparison against PySceneDetect on synthetic and real videos is a small but legitimate experiment, and the safe-zone algorithm is concrete. The statistics (scene counts, caption lengths, safe-zone area) give a reasonable picture of the data.\n\nThe soft spots are real and, for a dataset paper, central. The quality claim rests entirely on the author's own visual judgment. Section 2.1 says selection focuses on 'visual and aesthetic quality'; Section 2.4 says manual random sampling verifies quality. No inter-annotator agreement, no external human-preference assessment, no downstream fine-tuning comparison against Koala-36M or any other dataset. Figure 1 shows sample frames, but that is not evidence. So the headline advantage over Koala-36M is currently unsupported. The paper also leaves the motion filter threshold unspecified, and the dataset is not yet released, so the clips and captions can't be independently checked.\n\nNone of this is disqualifying. Many dataset papers ship with less. But the claim needs to be downgraded to 'manually selected with no quantitative validation' until the data is public and someone else can look at it. If a referee asked for one change, it would be: add a small fine-tuning experiment or a human preference comparison, and publish the exact thresholds.\n\nCitation patterns are fine; the self-citations are relevant and not padded. The paper is a single-author project and it shows some signs of rushing (a few grammatical slips, the missing threshold), but the thinking is coherent.\n\nRecommendation: send it to peer review, not desk reject. It deserves a serious referee, conditional on the data release and an external quality check. I'd bring it to reading group only if someone is actively working on video data; otherwise it's a maybe.","headline":"A genuinely new 170k-clip bilingual video dataset with a transparent pipeline, but the 'high visual quality' claim is self-validated and needs external evidence.","tokens_in":7703,"tokens_out":2594,"would_cite":false,"duration_ms":23307,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Tiger200K is a manually curated 170,000-clip bilingual video dataset built to give open text-to-video models a higher-quality fine-tuning resource.","keywords":["video dataset","text-to-video generation","user-generated content","data curation","bilingual captioning","shot boundary detection","safe zone","fine-tuning"],"falsifier":"A controlled fine-tuning experiment would settle the claim: fine-tune identical copies of one open video generation model on Tiger200K and on a matched random subset of Koala-36M with comparable clip counts and captions, then compare outputs on a fixed prompt suite through a human preference study or automated quality metric. If the Tiger200K-tuned model does not show measurably better visual quality or prompt adherence, the paper's quality claim is not supported. A second check is to compute an objective aesthetic score distribution on random clips from both datasets; if the distributions overlap heavily, the claimed quality gap disappears.","tokens_in":6757,"feed_emoji":"🎬","tokens_out":6296,"duration_ms":52640,"temperature":0.7,"pith_summary":"Tiger200K is built on a simple premise: for fine-tuning a video generation model, a small set of videos that humans have judged to look good beats a huge algorithmically filtered set. The paper claims that existing open datasets such as Koala-36M, though large, do not meet the visual quality bar for post-training or quality-tuning, and that careful human selection of creators and clips from UGC platforms can close that gap. The result offered is an open dataset of 170,000 fixed-length, temporally consistent clips from 4,151 source videos, each with Chinese and English captions plus safe-zone crop information. If the claim holds, the dataset gives the open-source community a directly usable resource for the supervised fine-tuning stage of text-to-video models.","feed_headline":"170K hand-picked clips aim to lift video model fine-tuning","feed_subtitle":"Human curation, not bigger filters, is the proposed fix for fine-tuning video generators.","key_machinery":"The load-bearing object is the manually selected source collection combined with the safe-zone computation. A safe zone is the region of a frame left after subtracting watermarks, subtitles, logos, and black borders, found by running PaddleOCR over every frame and scanning for persistent black regions, and clips whose safe zone falls below half the frame are discarded. Around this, the pipeline uses TransNetV2 for shot boundary detection with scenes shorter than 121 frames dropped, optical-flow-based motion filtering to remove static clips, and Qwen2.5-VL to produce dense bilingual captions. The paper's argument is that these mechanisms preserve temporal consistency and clean frames, while the manual curation at the front end supplies the aesthetic quality that algorithmic filtering misses.","core_discovery":"The central claim is that human expertise in data curation, applied at the input stage, is what separates data good enough for fine-tuning from data merely good enough for pretraining. The paper argues that UGC platforms now contain professionally made content, and that selecting top creators, searching by camera model and production keywords, and relying on recommendation systems yields videos whose visual and aesthetic quality exceeds what threshold-based algorithmic filtering of older web-scraped video can guarantee. The dataset then applies a pipeline of TransNetV2 shot detection, OCR and border-based safe-zone computation, optical-flow motion filtering, and Qwen2.5-VL bilingual captioning to turn those source videos into 85,000 scenes and 170,000 fixed 121-frame cuts. The intended contribution is a high-visual-quality, temporally consistent, bilingual video-text corpus for post-training and quality-tuning of video generation models.","pith_inferences":["A natural extension the paper does not run is a controlled fine-tuning comparison: train the same video model on Tiger200K and on an equally sized random sample of Koala-36M, then measure generation quality and prompt adherence on a held-out prompt set.","The subjective quality claim could be made measurable by collecting pairwise human preferences between clips from Tiger200K and clips from existing open datasets; the paper leaves that quantification to future work.","If the manual-curation approach transfers, the same platform-focused strategy could be applied to other regional UGC platforms to produce culturally diverse high-quality corpora rather than one platform's aesthetic.","One implicit consequence is that dataset curation effort may shift from building smarter automatic filters to building better creator-discovery and review workflows, since the paper locates the quality gain in human judgment at the input stage."],"forward_implications":["Open-source text-to-video models gain a fine-tuning set of 170,000 clips whose frames are cropped to overlay-free safe zones and whose captions support both Chinese and English prompt following.","Fine-tuning on Tiger200K should shift generation quality toward the polished, high-resolution look of professionally produced UGC rather than the average quality of large web-scraped corpora.","The per-clip bilingual captions make the dataset usable as a benchmark for caption quality and video-text alignment, not only as training material.","Because half the source videos are 4K or above, the dataset also supplies high-resolution material for video super-resolution and high-resolution generation research.","The pipeline itself is reusable: shot detection, safe-zone cropping, motion filtering, and VLM captioning can be applied to any new UGC source for continued dataset expansion."],"supporting_citations":[{"why":"Koala-36M is the algorithmic-filtering baseline that Tiger200K claims to surpass in visual quality for fine-tuning.","marker":"[24]"},{"why":"Panda-70M is the preceding large-scale captioning dataset whose transition detection and filtering Koala-36M improves on and Tiger200K inherits as context.","marker":"[7]"},{"why":"TransNetV2 supplies the shot boundary detection that defines scenes and drives the 121-frame cut segmentation.","marker":"[21]"},{"why":"PaddleOCR detects Chinese and English text for the safe-zone computation that removes subtitles and watermarks.","marker":"[13]"},{"why":"Qwen2.5-VL generates the dense bilingual captions that make the video-text pairs usable.","marker":"[3]"},{"why":"The DALL-E 3 captioning result justifies the choice of detailed textual descriptions for improving prompt following in generation models.","marker":"[4]"},{"why":"HunyuanVideo's data pipeline is cited as relying on expert manual annotation for SFT, supporting the premise that human curation matters.","marker":"[12]"},{"why":"Wan's pipeline similarly uses expert manual annotation, reinforcing the choice of manual curation for fine-tuning data.","marker":"[23]"}],"fun_headline_variants":["Tiger200K: human curation over filters for video AI","Hand-picked UGC clips aim to improve video model tuning","170K human-vetted video cuts for better fine-tuning","Manual curation beats algorithms for video training data","Tiger200K dataset: quality-first approach to video fine-tuning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the author's own judgment of 'visual and aesthetic quality' during manual selection and review is a reliable measure of the quality that matters for fine-tuning video generation models, since the paper never defines or measures that quality against an external benchmark.","fun_headline_variants_meta":{"raw":{"variants":["Tiger200K: human curation over filters for video AI","Hand-picked UGC clips aim to improve video model tuning","170K human-vetted video cuts for better fine-tuning","Manual curation beats algorithms for video training data","Tiger200K dataset: quality-first approach to video fine-tuning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000214,"raw_usage":{"total_tokens":1398,"prompt_tokens":890,"completion_tokens":508,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":426}},"tokens_in":506,"tokens_out":508,"duration_ms":5773,"temperature":1.0,"reasoning_tokens":426,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:30:10.760039+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled fine-tuning experiment would settle the claim: fine-tune identical copies of one open video generation model on Tiger200K and on a matched random subset of Koala-36M with comparable clip counts and captions, then compare outputs on a fixed prompt suite through a human preference study or automated quality metric. If the Tiger200K-tuned model does not show measurably better visual quality or prompt adherence, the paper's quality claim is not supported. A second check is to compute an objective aesthetic score distribution on random clips from both datasets; if the distributions overlap heavily, the claimed quality gap disappears.","supporting_citations":[{"cited_title":"Panda-70m: Captioning 70m videos with multiple cross-modality teachers","cited_arxiv_id":null,"evidence_quote":"Panda-70M is the preceding large-scale captioning dataset whose transition detection and filtering Koala-36M improves on and Tiger200K inherits as context."},{"cited_title":"Transnet v2: An effective deep network architecture for fast shot transition detection","cited_arxiv_id":null,"evidence_quote":"TransNetV2 supplies the shot boundary detection that defines scenes and drives the 121-frame cut segmentation."},{"cited_title":"Improving image generation with better captions","cited_arxiv_id":null,"evidence_quote":"The DALL-E 3 captioning result justifies the choice of detailed textual descriptions for improving prompt following in generation models."}],"review_version":1}