Repo: /home/user1/MythosMini — a 1,050,241-param byte-level (vocab 257) looped-shared-block transformer trained on fineweb-edu, benchmarked on WikiText-2 byte PPL, BLiMP, ARC-Easy against the Tiny-ML leaderboard (overall = mean of blimp, arc, normalized wiki score; efficiency = overall/log10(params+1) so PARAMETER GROWTH IS PENALIZED only logarithmically — doubling params costs ~5% efficiency but any overall gain counts fully). Current pipeline: procedural Dyck warmup -> 2B-token training (NAdamW, WSD schedule) -> 3-round recursive STM -> benchmark all checkpoints -> SLERP tournament -> SFT instruct phase -> instruct SLERP. Architecture: 4 prelude + 1 shared block looped 2-6x with per-loop LoRA + 4 coda, dim 96, RMSNorm/RoPE/SwiGLU, tied embeddings, DSA-style sparse attention (shared model-level lightning indexer, top-128 keys, 64 local window, FlexAttention short path, chunked long path, 1M max context), FP4 E2M1 QAT on all linears. Recent regression: moving the indexer from per-attention to one shared model-level indexer + detaching the indexer bias for prelude/coda blocks raised throughput 0.85->1.19 Mtok/s but the verification run scored 59.47 overall vs 60.40 before, AND its STM round 3 NaN-diverged (prior runs round 3 gave +0.57; NaN also seen once before with output-embedding centering). All-time best: 60.41 (recursive STM, commit accd37b). GPU: one RTX 5090 32GB, torch 2.10.0+cu128. Known: full-model torch.compile broken (inductor dtype bug + flex recompile issues), STM phase runs at only ~76 ktok/s vs training 1.19 Mtok/s. USER GOAL: push output quality as far as possible toward Gemma-class coherence (text in/text out) — ambition is the point; consider parameter growth, distillation, synthetic data, anything that fits one 5090.
Read /home/user1/MythosMini/src/mythosmini/model/transformer.py, src/mythosmini/training/loop.py, src/mythosmini/training/stm.py, src/mythosmini/training/optimizer.py, src/mythosmini/training/scheduler.py, src/mythosmini/config.py. Hunt for QUALITY bugs and theory problems: (1) is the shared model-level indexer (selection computed from embedding output, reused by all 9 blocks incl. prelude/coda with detached bias) theoretically weaker than per-layer indexers, and how to get quality back cheaply? (2) What plausibly causes STM round-3 NaN at constant lr 3e-4 on FP4+sparse weights (look at the detached-bias gradient path, FP4 STE, the indexer gate growth, missing eps/clamps)? Propose concrete fixes with exact code locations. (3) Any other quality-harming detail (init, lr, schedule, loop sampling, tied embedding scale).
