marcoshernanz/tfg-runs

TFG runs

The training runs of Marcos Hernanz's bachelor's thesis on Muon, the router and the bias balancer in Mixture-of-Experts models, as code/runs/ holds them in the thesis repository: every run's settings.json, metrics.jsonl and routing.jsonl, the sweeps' sweep.json, the tuned rates.json, the evaluated runs' specialization.json and annex.json, and the analyses' tables and figures. The checkpoints and early saves are not kept.

The paper Muon Makes the First Mixture-of-Experts Layer Harder to Balance (arXiv:2610.37914) is the short version of the thesis, and its results come from these runs. Its code is at github.com/marcoshernanz/muon-moe-routing.

Folder	What	Code
tuning/	The rate sweeps of AdamW, Muon and the momentum control and the β2 check, mechanism tier, 0.5B tokens, seed 0	tag frozen-2026-09-23
rq1/	Gate 1: all AdamW and all Muon, 1B tokens, seeds 0 to 2, with analysis/	tag frozen-2026-09-23
rq2/	The learning-rate stress grid: three partitions at 1× to 16× the tuned rates, seeds 0 to 2, the calibration runs and the balancer-speed control, with stress/	tag frozen-2026-09-23
pilot/	The geometry pilot: eight partitions at the tuned rates, 0.5B tokens, seed 0, three of them stopped at step 3,000	the working tree of 2026-09-24, no tag
rq3/	RQ3: eleven arms (the router's optimizer against the rest's, Muon on one part at a time, Muon at AdamW's rate, and both optimizers without the balancer), 1B tokens, seeds 3 to 5, with analysis/ (against all-AdamW) and analysis-muon/ (against all-Muon)	tag frozen-2026-09-25
confirmation/	The confirmation tier (108M active parameters): the rate re-check under recheck/ (AdamW and Muon at 0.125× to 1× of the tuned rates, 1B tokens, seed 0, with each sweep.json and the winners' rates.json), then all AdamW, all Muon and Muon with the router on AdamW at 0.7 times the winners, 3B tokens, seeds 0 to 2, with analysis/ (against all-AdamW) and analysis-muon/ (against all-Muon)	tag frozen-2026-09-26
balancer/	The balancer block: all AdamW and all Muon with the bias at its rate and four times faster, mechanism tier, 1B tokens, seeds 6 to 8, with analysis/	tag frozen-2026-09-26
balance_loss/	The balance-loss block: all AdamW, all Muon and Muon with the router on AdamW with DeepSeek-V3's sequence-wise balance loss at 1e-4 beside the bias, mechanism tier, 1B tokens, seeds 3 to 5, with analysis/ (each run read against its arm and seed in rq3/)	tag frozen-2026-09-28
offset-probe/	The offset probe: 12 runs that repeat the second chain's runs for their first 3,000 steps, seeds 3 to 5, each with one change (the first mixture's router input centred, the bias four times faster, the router on AdamW with Muon at 2e-3) or none, with each run's offsets.npz and probe.json	tag frozen-2026-09-25 with the probe's patch
scale-probe/	The scale probe: 10 runs that repeat the confirmation tier's runs for their first 3,000 steps: all Muon and Muon with the router on AdamW with the first mixture's router input centred at seeds 0 to 2, the three arms as run at seed 0, and all Muon with the bias four times faster at seed 0, with each run's readings.npz (per logged step and mixture: the bias, mean logit and score, load with and without the bias, row norms and lean, the offset, and six numbers on the scores' flat range and the bias's share of the choices) and probe.json	tag frozen-2026-09-26 with the probe's script

The first three are the program's first chain, run on 2026-09-23 and 24 on a host of four RTX 5090s (the thesis repository's context/decision/EVIDENCE-LOG.md section 16). The pilot ran on 2026-09-24 on one-card hosts (section 17). rq3/ is the second chain, run on 2026-09-25 and 26 on another host of four RTX 5090s (section 18). offset-probe/ ran on 2026-09-26 on a third such host (section 19, item 12), confirmation/ and balancer/ are the third chain, run on 2026-09-26 to 28 on a fourth (section 20), and scale-probe/ ran on 2026-09-28 on a host of two RTX 5090s (section 21, item 3). balance_loss/ ran on 2026-09-28 and 29 on a host of four RTX 5090s, the scale probe's (section 23). Its analysis pairs each run with RQ3's run of the same arm and seed, which it does not hold: after restoring, code/ai-experiments/balance_loss_2026_09/link.sh links those nine runs of rq3/ into it. The two probes are not blocks of code/runs/: the scripts that read them expect offset-probe/ in code/ai-experiments/rq3_reading_2026_09/offset_probe/outputs/mechanism/runs/ and scale-probe/ in code/ai-experiments/scale_probe_2026_09/outputs/runs/, so restore each there on its own (--include "offset-probe/**" or --include "scale-probe/**", then move the folder), and leave them out of the command below.

Restore into the repository with:

hf download marcoshernanz/tfg-runs --repo-type dataset --local-dir code/runs --exclude README.md --exclude "offset-probe/**" --exclude "scale-probe/**"

Citation

@misc{hernanz2026muon,
  title         = {Muon Makes the First Mixture-of-Experts Layer Harder to Balance},
  author        = {Hernanz, Marcos},
  year          = {2026},
  eprint        = {2610.37914},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2610.37914}
}