Reads SFT data from Space repo → Fine-tunes Llama 3.2-3B → Saves checkpoint
Training Config
loop_idx
lora_rank
epochs
learning_rate
max_seq_len
l4_diversity_w
5-Layer Architecture
L1
AssertionBreaker
Owned ground truth eval
G4
L2
RewardGuard
Gaming + Goodhart detection
G2
L3
MetaAudit
Editable outer loop + param self-edit
G1
L4
ArchiveDirector
Diversity-aware task sampling
G3
L5
HeuristicEvolver
Self-evolved sampling heuristic
G5
Loop Progress
Live Training Log
[ARIA]Dashboard initialized. Waiting for launch... [ARIA]Space repo: Aniket-2004/Meta-Finale [ARIA]SFT source: sft_data/sft_loop0.jsonl [INFO]All systems nominal. Click LAUNCH TRAINING to begin.