kinglacto/anlp-a2-training-logs

ANLP Assignment 2: training logs

Why this repository exists: all runs were logged to Weights & Biases, but W&B's free plan does not allow making projects public, so the W&B links cannot be opened by graders. Every scalar logged to W&B is republished here instead, unchanged:

• TensorBoard: open the Training metrics tab of this repository to browse every run's curves (Parts 1-2 are plotted against training tokens).
• CSV: csv/<part>/<run>.csv, one row per logged step with every metric.
• Config and final metrics: runs/<part>/<run>/config.json and summary.json.

Part	Run	Logged rows	Checkpoint
part1	dense-lr0.0005	194	-
part1	moe-4e-top2-lr0.001	195	kinglacto/anlp-a2-p1-moe-4e-top2
part1	dense-lr0.002	194	-
part1	dense-lr0.001	194	kinglacto/anlp-a2-p1-dense
part1	moe-1s3r-top1-lr0.001	195	kinglacto/anlp-a2-p1-moe-1shared-3routed-top1
part1	moe-4e-top1-lr0.001	195	kinglacto/anlp-a2-p1-moe-4e-top1
part1	moe-4e-top2-wide-lr0.001	195	kinglacto/anlp-a2-p1-moe-4e-top2-wide
part2	adamw-lr0.004	124	-
part2	adamw-lr0.002	124	kinglacto/anlp-a2-p2-adamw
part2	adamw-lr0.001	124	-
part2	adamw-lr0.0005	124	-
part2	muon-lr0.02	124	kinglacto/anlp-a2-p2-muon
part2	muon-lr0.005	124	-
part2	muon-lr0.01	124	-
part2	muon-lr0.04	124	-
part2	lion-lr0.001	124	-
part2	lion-lr0.0001	124	-
part2	lion-lr3e-05	124	-
part2	lion-lr0.0003	124	kinglacto/anlp-a2-p2-lion
part2	mars-lr0.001	124	-
part2	mars-lr0.008	124	-
part2	mars-lr0.004	124	-
part2	mars-lr0.002	124	kinglacto/anlp-a2-p2-mars
part2	sophia-lr0.0001	124	kinglacto/anlp-a2-p2-sophia
part2	sophia-lr0.001	124	-
part2	sophia-lr0.0003	124	-
part2	sophia-lr0.003	124	-
part2	sophia-lr3e-05	124	-
part2	lion-wd0.1-lr0.0001	124	-
part2	lion-wd0.1-lr0.0003	124	-
part2	sophia-lr5e-05	124	-
part3	decoding	13	-

Part 1: MoE vs dense FFN for vi/ja-en translation (incl. the dense learning-rate sweep). Part 2: optimizer learning-rate sweeps (24 runs). Part 3: decoding-strategy metrics (no training).