Sep 21 03:13 Clarify no-Nymeria bbox and joint QA counts 03:11 Remove NymeriaPlus from bbox train and val; include mapping_run in bbox-only recipe Sep 20 00:21 Upload folder using huggingface_hub Sep 8 02:34 Upload folder using huggingface_hub Sep 1 03:50 Mirror refined ScanNet and ScanNet++ QA release files 03:50 Add refined ScanNet and ScanNet++ QA image volume 030 03:48 Add refined ScanNet and ScanNet++ QA image volume 029 03:44 Refresh ScanNet and ScanNet++ QA metadata archive 03:39 Document refined ScanNet and ScanNet++ QA refresh Aug 28 16:16 Fix VSI video_frames bucket and add checkpoint grounding dashboard 13:20 Sync nested release metadata with root revision 13:20 Update training recipe: length buckets, object cap, fused LM loss Preserve the 128-view schedule; add fixed-N token-length bucketing, category detection <=20 policy, selective fused linear cross-entropy, FlashAttention/TileLang dependencies, refreshed counts and H200 OOM validation. Aug 26 09:08 Upload folder using huggingface_hub 09:07 Upload folder using huggingface_hub 07:53 Upload folder using huggingface_hub 07:52 Upload folder using huggingface_hub Aug 19 11:20 Add cost-aware dynamic micro-batching Batch N=1/2 at 4 samples per device, N=4 at 2, and N>=6 at 1 while preserving global batch 256 and all exact task quotas. Synchronize task/view-cost buckets across DDP ranks and document the N=128 peak-memory limit. 10:03 Sync nested VSI-590K recipe docs and checksums 10:03 Update VSI-590K training recipe, checksums, and docs 10:02 Sync nested metadata copy to VSI-590K recipe 10:00 Replace metadata with VSI-590K VQA and official Cambrian-P pose recipe 09:58 Add VSI-590K supplemental image volume 026 09:41 Add VSI-590K supplemental image volume 027 09:32 Add VSI-590K supplemental image volume 028 07:06 Upload folder using huggingface_hub 06:32 Filter joint detection QA below 0.10% target area Replace joint train/val metadata with camera-aware 0.10% best-view bbox-area filtering; update exact counts, quotas, manifests, trainer docs, and checksums. Image/model archives are unchanged. 00:00 Upload folder using huggingface_hub 00:00 Upload folder using huggingface_hub Aug 18 23:57 Upload folder using huggingface_hub 19:07 Upload folder using huggingface_hub 19:06 Upload folder using huggingface_hub 19:04 Upload folder using huggingface_hub 18:48 Upload folder using huggingface_hub 18:37 Upload folder using huggingface_hub 18:33 Upload folder using huggingface_hub 18:30 Upload folder using huggingface_hub 18:26 Upload folder using huggingface_hub Aug 15 10:35 Update detection semantics, training schedule, and checksums 08:36 Fix category-conditioned detection QA and joint pose sampling Aug 13 13:30 Publish complete QA recipe docs and missing ScanNet++ images Add 49 missing ScanNet++ images and update release docs/checksums for the 14-train/14-validation QA recipe. 13:30 Complete all single-view QA train and validation parts Normalized metadata with all four single-view detection/grounding parts, corrected sampling, and QA coverage audit. 12:09 Fix joint scene-normalized pose and box supervision Replace metadata/trainer so camera translation, box center, and box size share one fixed question-independent scene scale. Add validation shards, normalization manifests, fail-closed trainer checks, and corrected documentation. Existing image/model volumes are unchanged. 03:08 Add QA validation data and validation-enabled training code Aug 12 12:08 Upload folder using huggingface_hub 08:21 initial commit 
