JorgeAV
/

MR-JEPA

vision-language

Model card Files Files and versions

5.96 GB

Ctrl+K

Ctrl+K

1 contributor

History: 45 commits

JorgeAV's picture

Phase 4 final results with training analysis

9825ad6 verified 14 days ago

checkpoints
Phase 4 (SmolLM2 decoder) checkpoint - Stage 2 Epoch 3 14 days ago
mr_jepa
fix: target_encoder.py — respect config.jepa_loss_fn (smooth_l1/mse/cosine) instead of hardcoded MSE 21 days ago
results
Phase 4 final results with training analysis 14 days ago
.gitattributes

1.52 kB
initial commit 21 days ago
README.md

11.4 kB
fix: README.md — complete ablation table with all 13 experiments and CLI flags, add no_sigreg/vicreg_only 21 days ago
launch_ablations.py

2.81 kB
add: launch_ablations.py — complete ablation matrix with CLI commands for all 12 experiments 21 days ago
test_architecture.py

16.4 kB
fix: test_architecture.py — use os.path.dirname(__file__) instead of hardcoded /app for sys.path 21 days ago
train_mrjepa.py

39.8 kB
Fix TextEncoder.unfreeze_last: compatible with both AutoModel (Qwen3Model.layers) and ForCausalLM (model.model.layers) 21 days ago
train_phase2.py

45.7 kB
feat: add persistent Trackio Space for image logging (space_id + sync) 21 days ago
train_phase3.py

42.8 kB
Add complete Phase 3 training script with generative decoder + open-ended VQA 17 days ago
train_phase3_1.py

42.3 kB
Add Phase 3.1 training: gen_weight 2.0, gen_len 32, scheduled sampling, beam search 17 days ago
train_phase4.py

51.5 kB
Add Phase 4 training: SmolLM2-135M decoder + bridge MLP 15 days ago