Gromov loss checkpoints

Checkpoints for github.com/fbyrman/gromov: hyperbolic vision-language models trained on CC3M with GLIGEN box grounding, with the Gromov loss in place of the entailment cone. Each is 100,000 iterations at total_batch_size=512, stored as {"model", "iteration"} in fp32 with no optimizer state. manifest.json records each file's sha256.

checkpoint	model	supervision	config
power_inter_k3_vit_s/model.pth	PowerGromovInter, power_k=3	captions	configs/train_gromov_power_vit_s.py
power_inter_k3_vit_b/model.pth	PowerGromovInter, power_k=3	captions	configs/train_gromov_power_vit_b.py
power_all_vit_s/model.pth	PowerGromovAll, power_k_inter=2, power_k_intra=3	captions and boxes	configs/train_gromov_power_all_vit_s.py
power_all_vit_b/model.pth	PowerGromovAll, power_k_inter=2, power_k_intra=3	captions and boxes	configs/train_gromov_power_all_vit_b.py

Load and evaluate from a clone of the GitHub repo:

hf download freek23/gromov power_all_vit_b/model.pth --local-dir checkpoints
python scripts/evaluate.py --config configs/eval_zero_shot_classification_score_level.py \
    --checkpoint-path checkpoints/power_all_vit_b/model.pth \
    --train-config configs/train_gromov_power_all_vit_b.py \
    --results-file evaluate_score_level.json