autoresearch.jsonl

{"type":"config","name":"Qwen3.5-2B Japanese autotrain","metricName":"test_loss","metricUnit":"","bestDirection":"lower"}
{"run":1,"commit":"32c2f4a","metric":0,"metrics":{},"status":"crash","description":"baseline eval-only run crashed because autotrain.sh passed adapters/ to mlx_lm.lora before any adapter_config.json existed","timestamp":1773512652975,"segment":0}
{"run":2,"commit":"32c2f4a","metric":0,"metrics":{},"status":"crash","description":"baseline eval-only run still crashed because mlx_lm.lora defaults adapter_path to adapters unless passed as an explicit empty string","timestamp":1773512695052,"segment":0}
{"run":3,"commit":"e2bda52","metric":2.988,"metrics":{},"status":"keep","description":"baseline: base Qwen3.5-2B bf16 without adapters on fixed 800-example Japanese test split; also fixed eval-only adapter-path handling in autotrain.sh","timestamp":1773513162614,"segment":0}
{"run":4,"commit":"5483170","metric":2.495,"metrics":{},"status":"keep","description":"phase1 baseline lora: 4096-example balanced Japanese mix with 400 iters, lr=2e-5, num_layers=16, batch_size=2, grad_acc=8 cut test_loss from 2.988 to 2.495","timestamp":1773514060153,"segment":0}
{"run":5,"commit":"5483170","metric":0,"metrics":{},"status":"crash","description":"phase1 oasst-heavy mix (1024 dolly / 768 hh / 2304 oasst) crashed near iter 200 with Abort trap 6 after validation despite lower early train loss; likely MLX/runtime instability rather than dataset logic","timestamp":1773515239398,"segment":0}
{"run":6,"commit":"5483170","metric":2.497,"metrics":{},"status":"discard","description":"phase2 prompt prefix test: prepending a fixed Japanese-only instruction prefix to every prompt slightly worsened test_loss from 2.495 to 2.497","timestamp":1773516395555,"segment":0}