MindView
Collection
MindView resources for language-guided target-view generation: benchmark data, training data, and BAGEL-SFT weights. • 3 items • Updated
MindView-BAGEL-SFT
MindView-BAGEL-SFT is the final EMA checkpoint of BAGEL fine-tuned for language-guided target-view generation from multiple observations of an indoor scene. It accompanies MindView: Benchmarking Generative Multi-View Spatial Intelligence.
Code and instructions · Training data · Benchmark
Checkpoint
| Item | Specification |
|---|---|
| Base model | ByteDance-Seed/BAGEL-7B-MoT |
| Training data | 20,000 MindView-Train examples; six context RGB images and one target image per example |
| Artifact | Full fine-tuned EMA weights, ema.safetensors |
| Precision | BF16 |
| Size | 29,214,685,336 bytes (29.21 GB) |
| Release | v1.0 |
The checkpoint includes the model parameters used by the BAGEL runtime. Configurations, tokenizer files, and the VAE are obtained separately from the pinned base-model snapshot below. Training-resume state is excluded.
Download and use
From the MindView repository root, install huggingface_hub to obtain the hf
command, then download the release:
hf download wangyanbo/MindView-BAGEL-SFT ema.safetensors \
--revision v1.0 --local-dir models/MindView-BAGEL-SFT
hf download ByteDance-Seed/BAGEL-7B-MoT \
--revision 5019f57d168e5816e8f3f701b17cc816bb7cf24b \
--include "*.json" "*.txt" "ae.safetensors" \
--exclude "model.safetensors.index.json" \
--local-dir models/BAGEL-7B-MoT
Use the MindView runtime with its BAGEL dependencies installed. From the MindView repository root, after downloading and extracting MindView-Bench according to its guide:
unset CONFIG_FILE
export BAGEL_PYTHON=python
export BAGEL_MODEL_PATH="$PWD/models/BAGEL-7B-MoT"
export BAGEL_SFT_FINAL_CHECKPOINT="$PWD/models/MindView-BAGEL-SFT/ema.safetensors"
export INPUT_JSON="$PWD/MindView-Bench/data/test.jsonl"
export SCANNET_ROOT="$PWD/mindview-data/scannet"
export SCANNETPP_ROOT="$PWD/mindview-data/scannetpp"
export ARKITSCENES_ROOT="$PWD/mindview-data/arkitscenes"
bash evaluation/run.sh generate bagel-sft-final
BAGEL_PYTHON must select the prepared BAGEL environment. Alternatively, place
these absolute paths in config.env and set CONFIG_FILE explicitly. The same
checkpoint path is used by the project's MindCue workflow; see the
evaluation and
MindCue guides for full
configuration and GPU requirements.
Integrity and precision
SHA-256 of ema.safetensors:
504eb0958540b4873623e87eae8a971e9af545117ad0b4eb240e8400935143ac
Verify it with sha256sum models/MindView-BAGEL-SFT/ema.safetensors.
All 1,223 tensors were verified against the original FP32 checkpoint cast
directly to BF16. This matches the parameter precision selected by the project's
BAGEL inference loader. Export verification does not replace GPU inference;
generation outputs and benchmark scores were not rerun for this packaging step.
License and attribution
The BAGEL base model is released under Apache-2.0. Preserve its applicable copyright and license notices when using or redistributing this derivative checkpoint. MindView code and training data have separate terms; see the code license and MindView-Train data terms.