Phase 4 final results with training analysis

Upload results/phase4_smollm2.json with huggingface_hub

Phase 4 results - SmolLM2 decoder, 7 epochs

Phase 4 (SmolLM2 decoder) checkpoint - Stage 2 Epoch 3

Add Phase 4 training: SmolLM2-135M decoder + bridge MLP

Add Phase 3.1 training: gen_weight 2.0, gen_len 32, scheduled sampling, beam search

Upload checkpoints/hybrid_main_phase3_1_best.pt with huggingface_hub

Upload results/hybrid_main_phase3_1.json with huggingface_hub

Upload results/hybrid_main_phase3.json with huggingface_hub

Add Phase 3 training results

Add Phase 3 best checkpoint (epoch 0)

Add complete Phase 3 training script with generative decoder + open-ended VQA

Add Phase 3 training script (generative decoder + enriched evidence)

Upload checkpoints/hybrid_main_phase3_best.pt with huggingface_hub

Upload results/hybrid_main_phase3.json with huggingface_hub

feat: add persistent Trackio Space for image logging (space_id + sync)

Upload results/hybrid_main_phase2.json with huggingface_hub

Fix TextEncoder.unfreeze_last: compatible with both AutoModel (Qwen3Model.layers) and ForCausalLM (model.model.layers)

Phase 2: Add SOTA visual diagnostics (13 types) with Trackio image logging

Upload checkpoints/hybrid_main_best.pt with huggingface_hub

Upload results/hybrid_main.json with huggingface_hub

add: train_phase2.py — Phase 2 perception fine-tuning with 3-group differential LR optimizer

add: launch_ablations.py — complete ablation matrix with CLI commands for all 12 experiments

fix: test_architecture.py — use os.path.dirname(__file__) instead of hardcoded /app for sys.path

fix: target_encoder.py — respect config.jepa_loss_fn (smooth_l1/mse/cosine) instead of hardcoded MSE

fix: test_architecture.py — correct evidence_gate attribute check (gate_type='none' not gate=None), add dinov2 config test, compact formatting

fix: export get_dinov2_ablation_config from configs

fix: test_architecture.py — add comprehensive ablation tests (no_jepa, no_rollout, no_gate, K variants, loss_fn variants, SIGReg vs VICReg, purist config)

fix: README.md — complete ablation table with all 13 experiments and CLI flags, add no_sigreg/vicreg_only

fix: ARCHITECTURE.md — complete ablation table with all 13 experiments, CLI flags, dinov2/loss_fn/sigreg/vicreg ablations, footnote on no_rollout→no_jepa

fix: train_mrjepa.py — add VICReg loss, --no_sigreg/--use_vicreg CLI flags, log all ablation fields in results JSON, document no_rollout→no_jepa conflation

fix: mr_jepa.py — proper evidence gate disable via gate_type (no lambda hack), init ablation flags from config, update docstrings to current stack

fix: ablation.py — wire all disable flags, add missing experiments (dinov2, mse/cosine loss, no_sigreg, vicreg_only), add run()/load_results() methods, generate CLI commands

fix: backbones.py — DINOv3 via timm + Qwen3-Embedding-0.6B, add get_transform(), proper layer unfreezing for both architectures

fix: model_config.py — update defaults to DINOv3-L + Qwen3-Embedding, add jepa_loss_fn/use_jepa fields, add get_dinov2_ablation_config()

fix: SAM 3.1 → jetjodh/sam3.1 (non-gated mirror) in training script

fix: update ARCHITECTURE.md — SAM 3.1 → jetjodh/sam3.1, align with current component stack (DINOv3-L, Qwen3-Embedding)

fix: SAM 3.1 → jetjodh/sam3.1 (non-gated mirror)

Training script with DINOv3+Qwen3-Embedding, dtype fixes, all ablation flags

Update README with adapted component stack (DINOv3, Qwen3, SAM3.1, etc.)

Upload checkpoints/test_cpu_best.pt with huggingface_hub

Upload results/test_cpu.json with huggingface_hub

Remove __pycache__ files

Initial MR-JEPA codebase: architecture, training, evaluation, and tests

initial commit