yuanyuan10/train_3d

Qwen3.5 Cambrian-aligned 3D training data v0.1

This package contains the exact active QA/SFT shards and pose shards used by scripts/run_training.sh, a deduplicated RGB image tree, full-scene pose indexes for fixed-128 Cambrian-P pose-only sampling, and the Qwen3.5-2B/4B/9B base checkpoints. Python/CUDA environments are not included.

2026-08-13 geometry-scale correction

This revision replaces the previous metadata and trainer. The earlier package still supervised pose translation in meters while box QA followed a different scale convention, which was not valid for joint training. In this revision, camera translation, box center, and box size are all divided by the exact same fixed scene_scale_m. Do not mix an older metadata tar, trainer, or metric-label pose checkpoint with this revision. The 2026-08-10 metric-label overfit run only validated the pose gradient path; run the packaged normalized one-scene overfit before starting the formal 5K-step job.

2026-08-31 refined ScanNet/ScanNet++ QA refresh

After the screen-size and at-most-20-object gates, the joint-detection pool contains 1,172,041 train rows for N=2/4/6/8. ScanNet and ScanNet++ detection now use the postvalidated refined boxes. ScanNet visibility uses sensor depth; ScanNet++ visibility uses mesh-rendered depth aligned to the iPhone RGB cameras. Both sources also require projected short side >=8 px and area >=256 px^2 upstream. A joint row is retained only when every required answer object occupies at least 0.10% of a source image in its best input view. This camera-model-aware gate also uses the original observation bbox for retained non-ScanNet sources, including Nymeria FISHEYE624. N=2/4/6/8 train counts are 260,934/286,379/304,941/319,787. Every row has pose supervision and asks for a named category. These rows are combined with 288,662 official VSI rec spatial-VQA rows in the joint stream. The synchronized four-rank schedule visits the complete 1,460,703-row unique pool before its 550,101 alignment repeats. ScanNet/ScanNet++ grounding keeps the legacy text and box targets but removes targets failing the depth-visibility and projection-size checks. The full replacement and before/after distribution is in manifests/size_filter_report.json.

2026-08-19 fixed-128 pose-only alignment

Formal pose-only training directly uses the official preprocessed VSI scene pool: 2,899 ARKitScenes, 1,201 ScanNet, and 856 ScanNet++ scenes, spanning 5,105,870 pose-valid RGB frames. It samples uniformly over these 4,956 scenes. Pose-only training and pose validation use exactly 128 frames per sample, matching the released Cambrian-P main-model launcher. A shorter scene trajectory repeats its final selected frame to reach 128. The previous dynamic choice over 2/4/6/8/16/24/32/64/128 was our extension and is no longer used by the formal launcher. Joint category-detection rows keep their generated N=2/4/6/8 effective views and canonical anchor without runtime view padding.

2026-08-26 128-total-view QA batching

The formal four-GPU launcher keeps source QA images unchanged and targets about 128 aggregate views per local forward. N=2/4/6/8/128 use micro-batch 64/32/21-or-22/16/1, giving 128/128/126-or-132/128/128 total views. N=1 is capped at batch 64 by the existing 64-sample per-rank optimizer budget, which preserves global batch 256. Sample quotas, answers, and camera targets do not change. The exact forward count is derived from the refreshed per-view row distribution at runtime; backbone gradient checkpointing remains enabled.

2026-08-28 length-aware batching and fused language loss

Category detection remains single-category and is now capped at 20 answer instances in both package generation and the runtime loader. Within every fixed-N stream, a bounded 8,192-row look-ahead groups nearby text-token lengths; it does not create fake samples or change the 128-view schedule. The language loss applies the causal shift and selects only supervised assistant tokens before a fused linear cross-entropy LM head, so prompt, image and padding positions never allocate full-vocabulary fp32 logits. The formal launcher uses FlashAttention 2, FLA and TileLang.

2026-08-28 VSI video bucket and checkpoint visualization fix

The released VSI QA metadata names pre-extracted videos video_frames. The formal runtime now maps both video and video_frames to the N=128 cost bucket; the former predicate matched only video and failed when that stream was first scheduled. The packaged launcher also writes a live local dashboard and, at every 500-step checkpoint, greedily generates four fixed held-out grounding examples and projects predicted 3D cuboids into their input views. Run scripts/serve_training_dashboard.sh with the same OUTPUT_DIR to view loss, validation, padding, GPU memory, and the checkpoint gallery.

Layout and extraction

The release has one metadata tar, three model tars, and 31 independent image tar volumes. The legacy model-000.tar contains 9B; model-2b-000.tar and model-4b-000.tar contain the other two selectable backbones. Every tar has the same top-level directory qwen35_cambrian_3d_training_v0_1. Download all files into one directory, then run:

python -m pip install -U "huggingface_hub[hf_xet]"
hf download yuanyuan10/train_3d --repo-type dataset --local-dir ./train_3d
cd ./train_3d
chmod +x extract_and_verify.sh
./extract_and_verify.sh . /your/destination
cd /your/destination/qwen35_cambrian_3d_training_v0_1

Manual equivalent: verify with sha256sum -c SHA256SUMS, then extract the metadata tar, all three *-model-*.tar files, and every *-images-*.tar into the same destination. Do not concatenate the tar files.

Training

Install training/qwen3vl_3d/requirements-core.txt, then:

export CUDA_VISIBLE_DEVICES=0,1,2,3
# Choose 2B, 4B, or 9B. The backward-compatible default is 9B.
export MODEL_SIZE=4B
./scripts/run_training.sh

All three launchers accept MODEL_SIZE=2B, MODEL_SIZE=4B, or MODEL_SIZE=9B and resolve to the corresponding bundled checkpoint under models/. The backward-compatible default is 9B. Set MODEL_PATH=/another/checkpoint only when deliberately overriding the selected bundled model. Output directories include the selected size so switching models cannot silently overwrite a run.

The released Qwen directories remain standard Hugging Face checkpoints. The camera-query insertion, size-aware projector, Cambrian/VGGT camera head, and pose loss are supplied by the packaged Qwen35WithPose wrapper. A pose_head.pt is specific to the backbone hidden size, so do not reuse one across 2B, 4B, and 9B.

Before a long run, scripts/run_overfit_one_scene.sh provides a one-scene, single-GPU sanity/overfit test against the same packaged scene indexes. An optional multi-scene camera-head warm-up is available as scripts/run_pose_warmup.sh; pass its checkpoint to formal training with POSE_INIT=/path/to/pose_head.pt.

The exact camera-query insertion, pose target/loss conventions, training phases, and checkpoint limitations are recorded in MODEL_TRAINING_CHANGES.md. Read it before changing the model or resuming pose training.

The launcher defaults to a global batch size of 256 and 13,754 steps. Its exact global quotas are 640,000 pose-only samples, 2,010,804 joint language+pose samples (1,172,041 category-detection rows plus 288,662 VSI reconstruction spatial-VQA rows and 550,101 batch-alignment repeats), 568,032 existing QA-only samples, and 302,188 official VSI QA-only samples. All 1,460,703 unique joint rows are consumed without replacement across DDP ranks before 550,101 alignment repeats. The VSI QA-only stream consumes all 302,005 unique records before 183 batch-alignment repeats. Its 1,007 videos are pre-extracted once to fixed 128-frame JPEG sequences, so training does not repeatedly seek MP4 files. Existing QA-only shards are selected with exact row-count-proportional quotas per source N=1/2/4/6 bucket, then sampled in proportion to shard rows inside that bucket. VSI image N=1 and video N=128 retain separate accounting buckets and use local micro-batch 64 and 1, respectively. Validation is uniform across nine logical task families. Every 250 steps it evaluates both held-out camera pose and all 15 QA validation parts.

Detection always has a requested category. If exactly one matching instance is visible the row is detection_one; if multiple matching instances are visible the row is detection_all and returns all of that category. The obsolete all-categories/all-objects prompts are excluded from every model-facing shard. Category-detection answers with more than 20 instances are excluded as well.

All model-facing 3D geometry is dimensionless. A fixed, question-independent scene scale is shared by camera translation, box center, and box size. Rotations, intrinsics, and FoV are not scaled. The source metric scale is retained only in audit metadata and is never exposed in the prompt.

camera_translation_norm = camera_translation_m / scene_scale_m
box_center_norm         = box_center_m         / scene_scale_m
box_size_norm           = box_size_m           / scene_scale_m

Contents

• QA/SFT train: 17 selected JSONL shards, 6381799 rows.
• QA/SFT validation: 15 matched JSONL shards, 177039 rows.
• Formal pose-only: 4,956 official Cambrian-P/VSI scene identities.
• Official pose scene pools: 4,956 .npz indexes spanning 5,105,870 pose-valid frames; formal pose-only samples draw exactly 128 frames from these pools.
• Legacy pose validation: 703 scene-disjoint ScanNet++ rows.
• Images: 10,489,909 canonical files, deduplicated across pose and QA.
• Models: Qwen3.5-2B, Qwen3.5-4B, and Qwen3.5-9B BF16 checkpoints (about 4.3/8.7/18.0 GiB respectively), selectable with MODEL_SIZE.
• Model provenance: manifests/model_manifest.json records the exact upstream repository/revision, archive mapping, default size, and pose-head compatibility.
• Recipe: manifests/training_data_manifest.json records every train/validation shard, row count, and sampling weight.
• Geometry: manifests/scene_scales.jsonl records the fixed scale and provenance for every scene; manifests/normalization_report.json records its audit; manifests/training_data_manifest.json records the complete joint-pool recipe.
• Screen footprint: manifests/size_filter_report.json records exact train/val before/after distributions for the 0.10% target-area hard gate.
• Integrity: SHA256SUMS, manifests/archive_plan.json, and manifests/images.jsonl.gz.

All paths used for loading images are relative to this package root. Raw video containers are intentionally not needed after extraction.

License and access

No license is granted by this packaging process. Redistribute only the portions for which you have permission, and preserve the original dataset terms. In particular, ScanNet++ access is subject to its own terms and must not be publicly redistributed without the required authorization.