cyun9286/ABot-Recon-ft13-step2000

ABot-Recon: 13-source fine-tuning, step 2000

Fine-tuned from acvlab/ABot-Recon, using the ABot-Recon training branch with adapters for 13 data sources. This is the checkpoint at 2000 successful optimizer updates.

Files and loading

• abot_recon.safetensors: full EMA weights, recommended for inference and used for the reported validation.
• online/abot_recon.safetensors: full online model at the same step.
• training_summary.json: training settings, source batch sizes and validation results.
• loss_curve.png: training and validation loss through step 2000.

Install the ABot-Recon source package and its inference dependencies, then load the EMA export:

from abot_recon import ABotRecon

model = ABotRecon.from_pretrained(
    "cyun9286/ABot-Recon-ft13-step2000",
    filename="abot_recon.safetensors",
    loop_closure=False,
)

This model repository is publicly accessible. The exports contain model parameters. Optimizer, RNG and EMA update state remain in the original training checkpoint.

Training

• 16 NVIDIA H20-3e GPUs; bf16; AdamW; peak learning rate 1e-6.
• Encoder frozen; other non-confidence training parameters enabled; confidence fine-tuning disabled.
• EMA decay 0.999; 504 x 280 images; local attention window 12 frames.
• 32-frame clips and batch 4 per GPU for most sources; Kubric uses 16-frame clips and batch 8; Spring uses 32 frames and batch 1.
• Weighted sampling of BlinkVision indoor/outdoor, Virtual KITTI 2, Dynamic Replica, TartanAir, MVS-Synth, ScanNet, Spring, ScanNet++, PointOdyssey, Kubric, GTA-SfM and OmniWorld.
• Representative subset, not full-dataset training: 368 prepared items, 45,252 training frames and 5,192 validation frames. The held-out scene split is separate from training.

Validation

Validation uses EMA weights, up to four batches per source, and a mean across the 13 sources. These are training-objective losses on a limited held-out subset, not full benchmark accuracy scores.

Optimizer step	Validation loss
500	0.04847221
1000	0.04623022
1500	0.04416771
2000	0.04185896

Attribution and license

This fine-tuned model is derived from ABot-Recon, which builds on Pi3 weights. The weights retain CC BY-NC 4.0 and the upstream attribution and restrictions described in MODEL_LICENSE.md and NOTICE. Dataset licenses remain separate. This repository does not contain training data.