Stage 3 SFT best (step 3550, loss 3.6461) - pruned 80K vocab

Update model: SFT step 3550, loss=3.6461

Update model: SFT step 3000, loss=3.8530

Update model: SFT step 2950, loss=4.0059

Update model: SFT step 2650, loss=4.0579

Update model: SFT step 2100, loss=4.0962

Update model: SFT step 1825, loss=4.3889

Update model: SFT step 1525, loss=4.4235

Update model: SFT step 1050, loss=4.6154

Update model: SFT step 1025, loss=4.9765

Update model: SFT step 850, loss=5.0193

Update model: SFT step 675, loss=5.3019

Sync latest aetheris source code

Add Aya tokenizer files (avoid gated repo dependency)

Verified raw state_dict (weights_only=True compatible)

Update model card with full architecture and training details

Raw state_dict from SFT step 200

Remove vocab mapping (using full vocab SFT model)

Full vocab config for SFT model

Stage 3 SFT checkpoint (step ~115, in-progress training)

Upload pruned Aetheris (536M params, 80K vocab, 25.7% smaller)

Upload Aetheris model with source code (Stage 2, 722M params, loss=2.73)

Upload Aetheris model (Stage 2 best, 722M params, loss=2.73)

Update Stage 2 metadata: COMPLETE, best loss=2.7305

Upload Stage 2 final checkpoint (step 20000)

Upload final Stage 2 best checkpoint (loss=2.7305, 20K steps)

Stage 2 checkpoint: [Step 18500/20000] loss=3.1250

Stage 2 checkpoint: [Step 16250/20000] loss=3.1719

Stage 2 checkpoint: [Step 14050/20000] loss=3.7812

Stage 2 checkpoint: [Step 13450/20000] loss=3.3594

Stage 2 checkpoint: [Step 12350/20000] loss=3.3750

Stage 2 checkpoint: [Step 10150/20000] loss=3.4375

Stage 2 checkpoint: [Step 8500/20000] loss=3.0625

Stage 2 checkpoint: [Step 7350/20000] loss=3.0469

Stage 2 checkpoint: [Step 6250/20000] loss=4.3750

Stage 2 checkpoint: [Step 5150/20000] loss=4.0312

Stage 2 checkpoint: [Step 4550/20000] loss=3.7031

Stage 2 checkpoint: [Step 4000/20000] loss=4.1250

Stage 2 checkpoint: [Step 3400/20000] loss=3.6719

Stage 2 checkpoint: [Step 2850/20000] loss=4.2500

Stage 2 checkpoint: [Step 1800/20000] loss=5.8750

Stage 2 checkpoint: [Step 1200/20000] loss=6.2812

Stage 2 checkpoint: [Step 650/20000] loss=6.4688

Stage 2 checkpoint: [Step 100/20000] loss=7.5625

Stage 1 checkpoint: [Step 50/20000] loss=7.7500

Stage 1 complete: 10K steps, CKA layer alignment final checkpoint

Stage 1 checkpoint: [Step 9100/10000] loss=0.0141 cka_mean=0.1489

Stage 1 checkpoint: [Step 8400/10000] loss=0.0131 cka_mean=0.1052

Stage 1 checkpoint: [Step 7100/10000] loss=0.0161 cka_mean=0.1370

Stage 1 checkpoint: [Step 6400/10000] loss=0.0199 cka_mean=0.3752