Issues

franken (dense): extend per-layer FRANKEN_POLICY to the FFN/MLP weights — the leverage F1 lacked, where the literature says DFA actually works (Tao F14)

#151

· OriPekelman opened

on Aug 7, 2026

franken-moe: --dfa-feedback kolen-pollack — adaptive/learnable expert feedback that tracks the forward weights (Tao F13, the untested lever)

#150

· OriPekelman opened

on Aug 6, 2026

toy eval ce rejects franken-moe --ckpt-every checkpoints (model vocab -1) — blocks clean CPU held-out CE for the F-series

#149

· OriPekelman opened

on Aug 6, 2026

franken-moe: --lr-control reactive — a loss-reactive LR damper for block-DFA (the last LR lever, Tao F9M)

#148

· OriPekelman opened

on Aug 4, 2026

Per-layer routers at depth: top1 + --shape deep (toy#145 regression, fixed) — and what the aux loss now means

#147

· OriPekelman opened

on Aug 4, 2026

franken-moe: per-layer LR schedule (--lr-schedule ramp-up|ramp-down) — the next block-DFA lever

#146

· OriPekelman opened

on Aug 3, 2026

franken-moe --shape deep — the depth lever for block-DFA (F9f/F9g fallout)

#145

· OriPekelman opened

on Aug 2, 2026

Block-level DFA (--dfa-granularity block) — the literature recipe that actually works on transformers (F9d fallout)

#143

· OriPekelman opened

on Aug 1, 2026

K4: Stable LatentMoE — latent expert sandwich + shared experts (toy#135)

#142

· OriPekelman opened

on Aug 1, 2026

franken-moe: --freeze-experts — the frozen-expert control (R2, the inert-experts test)

#141

· OriPekelman opened

on Jul 31, 2026

franken-moe: --donor <gguf> embedding transfer — the F10 (transfer × DFA) surface [lower priority than muon #139]

#140

· OriPekelman opened

on Jul 30, 2026

franken/franken-moe: --optimizer adamw|muon|sgd — the F9r2 optimizer-sensitivity test (muon = engine work, sgd nearly free)

#139

· OriPekelman opened

on Jul 30, 2026

Navigation Menu

All issues

franken (dense): extend per-layer FRANKEN_POLICY to the FFN/MLP weights — the leverage F1 lacked, where the literature says DFA actually works (Tao F14)

franken-moe: --dfa-feedback kolen-pollack — adaptive/learnable expert feedback that tracks the forward weights (Tao F13, the untested lever)

toy eval ce rejects franken-moe --ckpt-every checkpoints (model vocab -1) — blocks clean CPU held-out CE for the F-series

franken-moe: --lr-control reactive — a loss-reactive LR damper for block-DFA (the last LR lever, Tao F9M)

Per-layer routers at depth: top1 + --shape deep (toy#145 regression, fixed) — and what the aux loss now means

franken-moe: per-layer LR schedule (--lr-schedule ramp-up|ramp-down) — the next block-DFA lever

franken-moe --shape deep — the depth lever for block-DFA (F9f/F9g fallout)

Block-level DFA (--dfa-granularity block) — the literature recipe that actually works on transformers (F9d fallout)

K4: Stable LatentMoE — latent expert sandwich + shared experts (toy#135)

franken-moe: --freeze-experts — the frozen-expert control (R2, the inert-experts test)

franken-moe: --donor <gguf> embedding transfer — the F10 (transfer × DFA) surface [lower priority than muon #139]

franken/franken-moe: --optimizer adamw|muon|sgd — the F9r2 optimizer-sensitivity test (muon = engine work, sgd nearly free)

Search results