docs: drop benchmark chart (redundant with table + stale ~15M caption)

docs: dedup — drop duplicated intro, trim ATB section

docs: add training-loss, benchmark, and token-budget figures

docs: results now from SFT-30K checkpoint (was stage-1 retrieval)

docs: professional model card + animated architecture GIF

Update model card: Phase 1 complete, Phase 2A in progress

Upload folder using huggingface_hub

Restore full README: merge original theory/citation/docs with current status

Update README: checkpoint-44000, add full eval metrics table

Update root model to checkpoint-44000

Add checkpoint-44000 (step 44k, 58.7% training)

Update README: checkpoint-37000 is now the default model

Update root model to checkpoint-37000 (step 37k, primary_loss=1.98, best so far)

Add saliency visualizations: saliency_viz/step_36k_cpu

Add saliency visualizations: saliency_viz/step_36k

Add saliency visualizations: saliency_viz/step_35k

Add saliency visualizations: saliency_viz/step_28k

Add checkpoint-37000 (step 37k, loss=4.06, primary=1.98)

Upload checkpoint-33000 (best loss=4.0855, step=33000)

Upload checkpoint-33000 (best loss=4.0855, step=33000)

Update README for checkpoint-30000

Update to checkpoint-30000 (step 30k, fixed saliency regularizer)

Update to checkpoint-30000 (step 30k, fixed saliency regularizer)

Update README to checkpoint-28000 (37.3% trained)

Update to checkpoint-28000 (step 28000/75000, loss 4.92)

Update to checkpoint-28000 (step 28000/75000, loss 4.92)

Add preprocessor_config.json

Add model card with usage instructions

Upload checkpoint-5000 (step 5000/75000)

initial commit