acceptodds
Under review as a conference paper at ICLR 2027

VidaForge: Open Research Infrastructure for Video Pretraining Data Recipes

Abstract

Video foundation models increasingly rely on large-scale pretraining data, yet the end-to-end data pipelines behind them remain largely closed and difficult to inspect or reuse. Researchers seeking to understand how video data recipes affect model pretraining often need to build substantial infrastructure before testing even a focused hypothesis. We present VidaForge, an open research infrastructure that represents a video data recipe as an executable five-stage workflow from raw videos to training datasets. A decision in this workflow can be varied to construct alternative datasets while preserving how every sample was produced. To demonstrate this research workflow, we compare data recipes with different coverage and quality in early from-scratch pretraining of Wan 2.1 and V-JEPA 2.1. Across both learning objectives, the broader-coverage recipe achieves the highest downstream benchmark scores, while loss-based evaluation favors different recipes. This study demonstrates how VidaForge connects data-recipe choices to downstream model performance. We further release VidaForge- 3M, containing 3.14 million scene-level clips totaling 6,475 hours, with fine-grained annotations and curation signals for video data-recipe research.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
video foundation models, video data curation, video data recipes, research infrastructure, data attribution

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_DuHqOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

VidaForge is an infrastructure for constructing and comparing video pretraining data recipes through ingestion, segmentation, selection, annotation, and model-specific packaging. It retains processing records, intermediate outputs, and rejected examples so that recipe variants can reuse prior computation and remain traceable. A four-recipe study performs 24 from-scratch runs across Wan 2.1-1.3B and V-JEPA 2-1B, comparing a larger pool used once with smaller subsets used twice under matched update and processed-clip budgets. The paper reports higher mean downstream scores for the broader pool, a mismatch between validation-loss and downstream rankings, competitive operator throughput, and a 3.14-million-clip annotated dataset release.

Critical strengths
  1. The infrastructure addresses the full path from data decisions to model evaluation. Separating saved quality scores from selection decisions and retaining rejected clips directly enables the Selected, Rejected, Mixed, and Random comparisons. Appendix B specifies identifiers, error records, and stage dependencies rather than presenting traceability only as a high-level aspiration. Supporting both generative and self-supervised training gives the workflow practical breadth.

  2. The recipe study includes meaningful controls: a random subset of the same size, parent-video isolation from pretraining validation sets, three runs per recipe, and matched seeds for Mixed, Selected, and Rejected. Mixed exceeds Selected in all three matched runs for both VBench Total and SSv2. The reported means are 60.64 versus 59.74 on VBench and 16.256% versus 15.712% on SSv2. These results provide useful evidence within the explicitly limited early-pretraining regime.

Critical weaknesses
  1. The generation study provides limited evidence about useful video-generation quality. Section 4.2 acknowledges that the models do not yet consistently generate coherent videos. Table 13 reports very low scores on several semantic dimensions, including Human Action of 1.56 and Multiple Objects of 0.14 for Mixed, while Mixed is not best on Overall Consistency or Temporal Style. The removal-of-Dynamic-Degree analysis is helpful, but does not establish that the roughly 0.9-point Total improvement over Selected reflects a perceptually meaningful gain. The scope should remain an early-training metric comparison, rather than a general recommendation about final data quality or mature generation models.

  2. The released annotation resource is not semantically validated. Appendix E.1 defines annotation completeness through successful operations and schema checks, while Appendix I explicitly states that annotation accuracy has not been systematically assessed by human evaluation. Camera labels are also supplied to caption generation, so annotation errors could propagate across fields. Since fine-grained annotations are a major part of the dataset contribution, valid structure and full coverage do not establish their reliability for studying caption or camera-motion recipes. This leaves an important part of the resource's claimed research utility unverified.

Questions for authors
  1. Clarification: What do the existing generated examples and per-dimension results support about the practical meaning of the VBench difference? Can the conclusion explicitly distinguish a higher early-stage aggregate score from improved coherent, prompt-aligned generation?

  2. Clarification: Is there any existing semantic audit of camera labels, captions, or tags beyond schema validation? If none is available, which fields and recipe experiments should be treated as unvalidated in the release documentation?

4 comments
AuthorsOpenReview ↗

Response to Reviewer DuHq (1/2)

Thank you for carefully examining our design and experimental controls, and for identifying where further evidence would strengthen the paper.

W2 / Q2: Annotation quality and Camera input to captions

A:Our semantic audits support the quality of Camera, Caption, and Tag annotations. Tests on CAPability [1], a benchmark for caption correctness and coverage, further show that Camera metadata improves both measures. The full audit results are available in our anonymous repository (https://anonymous.4open.science/r/anno_ff-F4D8)

We audit clips sampled from VidaForge-3M, following the annotation schemas (Appx. C.1–C.3) and generation prompts (Appx. D.1–D.3). We selected 1,000 clips to cover all 78 label values across Camera categories and audited both Camera and Caption on these clips. Since Camera labels guide caption generation, we check both annotations on the same clips. We selected another 1,000 clips to cover all 65 Tag categories. As stated in the paper, Camera uses Gemma-4-E4B-it, while Caption and Tag use Qwen3.6-27B-FP8.

We first tried training reviewers without a video-pipeline background, but the camera-motion and caption-level rules required substantial training. An author familiar with the project therefore reviewed 100 clips per annotation type, and GPT-6-Astra reviewed the other 900. Reviewers checked annotations against timestamped thumbnails of the full clip, viewed in time order.

For Camera and Tag, we checked whether each existing label matched the visual content. $\text{Label support} = \text{supported labels} / \text{all labels} $ with a definite judgment. About 2.4% of Camera labels could not be judged and were excluded from the denominator. Since subjects and actions are multi-label Tag fields, we also checked for missing main subjects or actions.

Our captions have four levels of detail, L0–L3. We audit L3, the most detailed level. We follow CAPability’s [1] prompts and scoring rules for its four video dimensions: Action, Event, Camera, and Count. DeepSeek-V4.1-Flash drafts rubric questions, such as what action occurs or how the camera moves, and reference answers for each clip. GPT-6-Astra checks these references against the video thumbnails, excluding incorrect or uncertain items. This gives 6,558 valid questions for 1,000 clips (6.56 per clip).

Let �1��1�, �2��2�, and �3��3� count correct, incorrect, and omitted items under the task rules. Precision (PP) measures correctness, and Coverage (CC) measures how much target content is correctly described:

$

P=N+1N+1+N−1,C=N+1N+1+N−1+N0.P=\frac{N_{+1}}{N_{+1}+N_{-1}},\qquad C=\frac{N_{+1}}{N_{+1}+N_{-1}+N_0}.

$

We average the four dimension scores equally and report percentages.

Measure100 clips: author review1,000 clips: 100 author + 900 GPT-6-Astra
Camera label support73.90%76.33%
Tag label support93.27%95.55%
L3 Caption Precision90.83%89.82%
L3 Caption Coverage81.29%79.35%

Both reviews show high Tag support and high L3 Precision and Coverage. In most clips (74%), the multi-label Tag fields cover the main subjects and actions.

We evaluate our multi-level caption schema and Camera input on all 2,192 CAPability videos using its official questions and reference answers. As described in the paper, L0 summarizes subjects, actions, and scenes; L1 adds event order; L2 adds temporal changes, relations, and background; L3 adds detail and camera descriptions. We generate Camera tags with our schema, then use them as input for Qwen to write L0–L3 captions. GPT-6.1-Terra scores them with the official evaluation prompts.

On CAPability, higher caption levels improve both correctness and coverage, supporting our multi-level design.

Caption levelPrecisionCoverage
L047.35%18.07%
L147.49%33.28%
L250.38%37.99%
L351.97%41.19%

We use the same Qwen model and multi-level prompt, changing only Camera input. L3 results are:

CAPability L3 measureWithout CameraWith Camera
Camera-description Precision38.11%47.21%
Camera-description Coverage34.81%46.78%
Overall Precision49.73%51.97%
Overall Coverage38.11%41.19%

Camera input improves both camera-description metrics and the four-dimension averages.

The repository (https://anonymous.4open.science/r/anno_ff-F4D8) also compares Gemma and Qwen Camera labels and reports per-dimension caption scores with and without Camera input on the 1,000 VidaForge-3M clips. It includes a GPT-6-Luna baseline using the same CAPability prompts.

[1] Liu, Zhihang, et al. “CAPability: A Comprehensive Visual Caption Benchmark for Evaluating Both Correctness and Thoroughness.” NeurIPS, 2025.

AuthorsOpenReview ↗

Response to Reviewer DuHq (2/2)

W1 / Q1: What the early VBench difference means

A: Thank you for asking whether the roughly 0.9 gain in VBench Total reflects a visible improvement. As noted in Sec. 4.2, the models do not yet consistently generate coherent videos that match the prompts. We therefore checked whether higher VBench scores also mean better visual quality at this early stage.

Fig. 16 and Tab. 13 show the scores for each dimension. Using the official weights, Imaging Quality and Aesthetic Quality account for about 52% of the 0.9 gain. Subject Consistency, Background Consistency, Temporal Flickering, and Motion Smoothness account for another 36%. We focused our visual checks on these six dimensions.

For each dimension, we selected two groups of videos with large score gaps: one favoring Mixed and one favoring Selected, giving 12 groups in total. Each group compares videos generated by Mixed, Selected, and Rejected from the same prompt. One author watched the original videos. DeepSeek V4.1 Flash, GPT-6.1-Terra, and GPT-6-Astra each reviewed image grids containing every frame, with anonymous video labels and no VBench scores. Each review compared the videos on the relevant dimension.

In 11/12 groups, the author preferred the higher-scoring video or saw no clear difference: 6 groups favored the higher-scoring video and 5 were tied. Dskv4.1, Terra, and Astra gave the same direction or a tie in 10/12, 8/12, and 12/12 groups, respectively; they strictly preferred the higher-scoring video in 6, 6, and 5 groups. These reviews support a visual basis for the early VBench gap: the inspected examples show visible differences in image quality and temporal stability that help explain the roughly 0.9 gain.

The video comparisons and human and model judgments are available in our anonymous repository (https://anonymous.4open.science/r/anno_ff-F4D8).


On a personal note, it meant a great deal to me that you took the time to examine the details in the appendix and understand the reasoning behind the design choices. Much of the work behind research infrastructure can be easy to overlook once a pipeline runs. Seeing that work understood and its limitations discussed with equal care was genuinely encouraging. Thank you.

Thank you for your response. I will maintain my score.

AuthorsOpenReview ↗

Many thanks for your response!

Thank you for getting back to us and for the care you put into reviewing our work.

Reviewer_byyhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduce a video pretraining data pipeline, which contains 5 stages of processing.

Critical strengths
  1. The paper present a data processing work for video foundation model training. This fill the gap that most existing works don't specify their data processing
  2. This work provides a data processing system that can be customized, and reproduced data processing procedures.
Critical weaknesses
  1. The processing itself lacks novelty. The paper presents a solid work requires a lot of effort, and it's helpful to community. However, its limited novelty doesn't fit as an ICLR paper. It basically uses existing data curation techniques.
  2. The experiments are limited to early pre-training stage. We should be careful to demonstrate that the selected data is better, since the training dynamics changes [1], and the lead on early stage training cannot guarantee the ultimate performance.
  3. Lack of ablations. The ablation on filters and thresholds are absent, it is reckless to conclude that the propose pipeline works.

[1] Pruthi, Garima, Frederick Liu, Satyen Kale, and Mukund Sundararajan. "Estimating training data influence by tracing gradient descent." NeurIPS 2020.

Questions for authors

Could the authors provide the end-to-end computational cost? Particularly, the efficiency gains from reusing intermediate results would be interesting.

4 comments
AuthorsOpenReview ↗

Response to Reviewer byyh (1/2)

Thank you for recognizing the effort behind VidaForge and its value to the community!

Q1: Computational cost and reuse of intermediate results

A: Appx. E.2 lists the time, throughput, and hardware used at each step to process 800k source videos into about 3.14M annotated clips. We summarize Stages I–IV below. CPU steps use Intel Xeon Gold 6530 processors, and GPU steps use NVIDIA H200 GPUs.

StageCPU core-hoursH200 GPU-hours
I. Ingestion4,008—
II. Segmentation14,533—
III. Selection13,134257
IV. Annotation—1,335
Total31,6751,592

Using Google Cloud's public on-demand prices as a reference, $$0.048433/CPU core-hour and $10.60/H200 GPU-hour, we estimate a compute cost of about $18,400.

This compute cost may be modest for frontier labs such as Meta MSL or DeepMind, but it is close to what an academic lab like ours can afford for data construction alone. On top of this, we completed 24 from-scratch pretraining runs: four data recipes, three runs each, across two model families.

For academic researchers like us, can this investment support further video data recipe experiments? VidaForge-3M retains the processed videos, scores, and multi-level annotations. Researchers can change existing filter thresholds or select and mix L0–L3 captions without rerunning the scoring or annotation models:

Recipe changeResults reusedAdditional scoring / annotation cost
Enable, disable, or adjust existing filtersProcessed videos, scores, and annotations$0
Change the choice or mixture of L0–L3The same videos, all four caption levels, and other annotations$0

Our pretraining experiments already reuse these processed videos, scores, and annotations.

Recent visual generation reports offer two examples of research questions this reuse supports.

1)The MOSS-Image report (October 2026) [1] found that common quality filters lowered several image-generation scores. It therefore used a broader range of data for final pretraining. Does this also hold for video pretraining? How do common aesthetic and text filters affect video model learning?

Technical reports cover many parts of model development. A focused academic study can examine when these filters help and why their effects vary across video content and training stages. Researchers can use VidaForge-3M's existing scores to vary selection conditions while reusing the videos and annotations for controlled comparisons.

2)The level of detail in captions is another example. MOSS-Image mixes short, medium, and long descriptions during pretraining. Krea 3 [2] (November 2026) found that long prompts alone worked better than a mix of long and short prompts in its setting.

Even frontier labs adopt different caption training recipes for visual foundation models. How much caption detail helps video pretraining, and how should different levels be mixed? VidaForge provides four caption levels, L0–L3, from brief summaries to detailed descriptions. Researchers can keep the videos fixed and vary the caption levels and their mixing ratios.

Researchers can use coding agents to add a new filter, compute its scores, and reuse existing videos and annotations. Those working with their own data can build on the same infrastructure. It has supported million-clip processing and training two model families, saving researchers the work of connecting and testing the pipeline from scratch.

We share VidaForge’s data infra and dataset so academic groups can build on each other’s work. Reusing VidaForge leaves more time and budget to study how video data affects model learning.

[1] MOSI.AI. MOSS-Image Technical Report. October 2026.

[2] Millon, Erwann, et al. Krea 3 Technical Report. November 2026.

AuthorsOpenReview ↗

Response to Reviewer byyh (2/2)

W1: Contribution beyond existing processing techniques

A: We agree that VidaForge uses existing data curation methods. However, its novelty also lies in how it supports video data research and the experiments built on it.

Using VidaForge, we study a practical question in video pretraining: how do data curation choices affect model learning under a fixed training budget? We test this through controlled pretraining experiments with generative and self-supervised models.

We also test VidaForge’s use of structured Camera tags in caption generation by comparing captions with and without these tags. On CAPability, a benchmark for caption correctness and coverage, Camera input improves both measures. See §9.1 (CAPability) of the full annotation quality evaluation: https://anonymous.4open.science/r/anno_ff-F4D8 for the setup and results.

Our L0–L3 captions let researchers vary caption detail and mixing ratios while keeping the videos fixed. Later reports from MOSS-Image [1] and Krea 3 [2] use different caption recipes. Focused academic studies can test when these choices help and how caption levels should be mixed across training settings. VidaForge provides the captions and training workflow needed to study these questions.

Researchers can build on VidaForge’s data infra and dataset to study diverse video data recipes. We believe this contribution is valuable to the ICLR community.

W2: Early training and changing recipe rankings

A: Thank you for raising this important question. We fully agree that an early lead does not guarantee better performance after longer training.

Figs. 18–19 in Appx. G.4 show this change: in both Wan and V-JEPA, Selected leads Mixed at the first downstream evaluation, but Mixed later overtakes it and finishes with higher mean VBench Total and SSv2 accuracy. Looking only at the first checkpoint would give a different ranking. We also find that loss rankings on shared validation sets differ from downstream rankings (Secs. 4.2–4.3).

As stated in Sec. 4, we conduct these early-pretraining experiments within a compute budget that we can afford, comparing recipes with the same training budget. The Mixed advantage we report therefore applies to the video pool, models, and training budget we tested. Your point also applies to this result: the relative performance of the recipes may change as training continues.

W3: Filter and threshold ablations

A: Thank you for the suggestion. Our 4 recipes × 3 runs provide recipe-level ablations by comparing Mixed, Selected, Rejected, and Random under a fixed early-pretraining budget, testing how VidaForge supports controlled research on video data recipes (Sec. 4.1, lines 265–267).

Before training, we inspected videos through our viewer and set the quality-filter thresholds based on visible quality issues. We then held these settings fixed for the comparisons. Using VidaForge, we built these dataset variants and trained and evaluated both generative and self-supervised models. These completed experiments demonstrate that VidaForge can turn data-selection questions into controlled training studies.

The individual filter and threshold ablations you suggest would further test how sensitive the training results are to the selection settings. These ablations would deepen our understanding of specific selection settings. VidaForge also supports these experiments. Researchers can use the saved scores and retained rejected videos to disable a filter or change a threshold, creating new training recipes without repeating scoring or annotation.

I appreciate the authors' detailed rebuttal and add experiments. However, my primary concern of limited novelty and using existing data curation methods are not addressed. Overall, I maintain my rating at 2 (Weak Rejection).

AuthorsOpenReview ↗

Many thanks for your response!

Thank you for considering our response. We respect your research taste, and we would like to make the research taste behind this paper clearer.

We believe a paper can contribute a new research perspective without introducing a new algorithm or curation method. We still do not fully understand how existing curation methods affect video pretraining. Even frontier labs use different recipes. Studying these questions should not require each lab to build an entire processing pipeline first.

We built VidaForge around this goal and demonstrated that it supports controlled studies of video data recipes through model pretraining.

We believe this is a research contribution in its own right.

Reviewer_U253OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces VidaForge, an open-source research infrastructure for process and study pretraining data pipelines for video foundation models. The authors have pointed out that such data pipelines are usually closed-sourced, and research community need to rebuild the infra to just test the basic data hypothesis. VidaForge tends to solve this issue by converting video data recipes into executable workflow.

The authors then validated by pretraining 2 models (Wan2.1 and V-JEPA2.1). The paper discovered that data coverage gives better downstream performance. More importantly the paper also released opensourced dataset with 3.14 million annotated video clips.

Critical strengths
  1. VidaForge provides everyone raw metrics and metadata, and let the community to set the own recipes on filtering. This is flexible for the researcher and define and test the own sampling strategies.

  2. This proposed framework is built on robust infrastructure.

  3. The 3.14 million clips the authors released are hightly curated, and with different meta data on different levels of captioning, camera labels, tags, and properties.

Critical weaknesses

W1. The released dataset (VidaForge-3m) is strictly engligh, and the scale is relatively small. This limits the utility for training multilingual video models. The authors downsampled high-res videos, resulting in a heavily skewed distribution on resolution.

W2. The dataset's souce is from llava-onevision-2 data, the fact that the source is homogenious limit the utility of the data, especially for pretraining stage.

W3. wonder what are the differences between vidaforge and nvidia's NeMo / Flora?

Questions for authors

please refer to the weakness session part. Thanks!

3 comments
AuthorsOpenReview ↗

Response to Reviewer U253

Thank you for recognizing the value of releasing raw measurements and configurable recipes!

W1: English annotations, dataset scale, and resolution

Concern: English-only annotations, limited scale, and downsampling may restrict the dataset's usefulness.

A: Thank you for raising these questions about dataset scale, language, and resolution.

VidaForge-3M contains 3.14M video clips totaling 6,475 hours, about three times the clip count and duration of OpenVid-1M [1], a dedicated dataset work. VidaForge-3M is one part of our contribution. Together with our open data infra, it supports research on video data recipes. Each clip includes four caption levels, Camera and semantic tags, quality scores, and duplicate relations. Researchers can reuse these results to build and compare different video data recipes.

The current captions are in English. We plan to translate all four caption levels into more languages to support multilingual text supervision, without processing the videos again.

VidaForge downsamples videos only if their shorter side is over 720 pixels. All source videos used for VidaForge-3M had a shorter side of 720 pixels or less, so none were downsampled. The clips keep their source resolution. The concentration at certain resolutions comes from the input data. We will clarify this in the paper.

W2: A single source and content diversity

Concern: Using LLaVA-OneVision-2-Data may yield homogeneous pretraining data.

A: LLaVA-OneVision-2-Data [2] is a large video corpus for multimodal model mid-training. It brings together several public video datasets. The videos included in VidaForge-3M originate from HowTo100M [3], Sekai [4], and PE-Video [5], accessed through this collection. They cover a range of activities and scenes.

The Tag distributions in our paper also show that VidaForge-3M covers indoor activities, food, products, people, cities, and nature. We release these tags so researchers can study the content distribution and adjust category proportions to build different pretraining data recipes.

[1] Nan, Kepan, et al. “OpenVid-1M: A Large-Scale High-Quality Dataset for Text-to-Video Generation.” ICLR, 2025.

[2] An, Xiang, et al. “LLaVA-OneVision-2: Towards Next-Generation Perceptual Intelligence.” arXiv:2605.25979, 2026.

[3] Miech, Antoine, et al. “HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips.” ICCV, 2019.

[4] Li, Zhen, et al. “Sekai: A Video Dataset towards World Exploration.” arXiv:2506.15675, 2025.

[5] Bolya, Daniel, et al. “Perception Encoder: The Best Visual Embeddings Are Not at the Output of the Network.” arXiv:2504.13181, 2025.

AuthorsOpenReview ↗

Response to Reviewer U253-Part II

W3 / Q: Differences from Nvidia NeMo and FLORA

A: NeMo is a broad ecosystem covering data processing, model training, and inference. Its most directly relevant component is NeMo Curator. Table 2 and Appx. A compare its design with VidaForge. VidaForge treats data-building decisions as experimental variables that researchers can test through model training, reusing existing processing results to construct and compare recipes. Frankly, engineers can use coding agents to add these capabilities to other frameworks. The design question is which research workflows the infrastructure makes straightforward to carry out. Fig. 7 shows that our academic team’s framework delivers processing speed comparable to or better than industrial frameworks such as Alibaba’s Data-Juicer and NVIDIA’s Cosmos Curator across the tested workloads.

What do we want to study with such a framework? In NeMo’s video foundation model paper, the NeMo Curator [6] section focuses on large-scale video processing throughput, using GPU acceleration and workload balancing across stages. VidaForge asks how video training data should be built and how these decisions affect model learning. We study a concrete question: under a fixed training budget, is it better to use a broader video pool or repeat a smaller, filtered subset? We test this through generative and self-supervised pretraining.

We hope VidaForge can serve as an open anchor for video data research. Academic groups can use our data infra and public dataset to spend their limited budgets testing their own hypotheses. Their findings can guide industrial frameworks to better support these experiments and build shared knowledge about video training data.

FLORA’s TempSeg [7] and Captioner [8] models (October, 2026) work together to split robot and ego videos at subtask boundaries and generate per-chunk action captions. VidaForge builds a large-scale pipeline for general videos to support video data recipe research. The two serve different domains and goals. These FLORA models could also be added to VidaForge for segmentation and captioning in the robotics domain.

As it happens, I’ve recently been using FLORA extensively for chunk-level captioning of robotics videos. Happy to discuss!

[6] NVIDIA. Training Video Foundation Models with NVIDIA NeMo. 2025.

[7] FLORA-TempSeg: https://huggingface.co/nvidia/Flora-TempSeg

[8] FLORA-Captioner: https://huggingface.co/nvidia/Flora-Captioner

AuthorsOpenReview ↗

Update on multilingual captions

A quick update on language coverage: we’re using a machine translation model to translate all four-level caption levels in VidaForge-3M into Chinese and Spanish. We expect to finish before the rebuttal period ends. We’d welcome your thoughts on our response and would be happy to clarify any remaining points.

Update: We’ve completed the Chinese and Spanish translations for all four caption levels in VidaForge-3M. In this anonymous repo (https://anonymous.4open.science/r/anno_ff2-FF09/) shows five examples with the original English captions and both translations. It also includes a downloadable Parquet file with captions for 100 clips sampled at random from one translated shard.

Thank you for raising this point !

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.