Optimizer Replay Audit for Transformer-Like Training Traces¶
- Task ID:
computer_science.muon_optimizer_replay_audit - Domain:
computer_science - Subdomain:
machine_learning_optimization - Status:
test - Tags:
optimizer_replay,transformer_training,matrix_updates,numerical_stability,deep_learning_numerics,training_diagnostics
Public Summary¶
This page is generated from task metadata and selected public-safe excerpts.
Example B1 Prompt Excerpt¶
# Optimizer Replay Audit for Transformer-Like Training Traces
You are given a deterministic optimizer replay package in `data/`. The package contains initial parameters, saved gradient traces, a layer manifest, public replay settings, and validation-proxy probes.
Implement the replay path used by a Muon-style optimizer for transformer-like training:
1. Read `data/param_init.npz`, `data/grad_trace.npz`, `data/layer_manifest.csv`, `data/replay_config.json`, and `data/validation_probe.npz`.
2. Maintain one momentum buffer per parameter.
3. Use the public learning-rate schedule:
- linear warmup for `warmup_steps`;
- then linear decay from `base_lr` to `base_lr * final_lr_fraction`.
Notes¶
- This page is a generated site artifact.
- Higher-level prompt details and internal benchmark specifics may remain intentionally undisclosed.