You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[IGLA-Coder] P0 Decision and honesty foundation #1033
Part of epic #1032 (IGLA-Coder). Plan: t27_coder_model_plan.md P0.
Goal
Lock the decision metric, freeze the baseline, and write the falsification-ledger row before any training. No compute.
Sub-tasks
P0.1 Write docs/IGLA_CODER_METRICS.md: BPB on held-out code corpus (necessary) + MultiPL-E pass@1 (sufficient). HumanEval/MBPP are Python-only and do not satisfy the "any language" requirement.
P0.2 Freeze baseline: record champion BPB=2.2111 provenance and that it is the standard AdamW arm (beta1=0.9, wd=0.04), NOT phi-canonical. SHA-256 the frozen config.
P0.3 Commit an 80/20 blind split seed pre-search (SPLIT_SEED=17 convention).
P0.4 Ledger row: "phi-anchored HPs help code BPB/pass@1" = [Open conjecture]; falsified if a non-phi control of equal tuning budget matches within Monte-Carlo error.
Compute decision applied: P0-P2 run at $0 (local/sandbox). Start P0 now - no GPU needed. Railway = control plane + Postgres ledger + artefact storage; GPU = rented worker only from P3 onward. See epic #1032 compute-decision comment.
CORRECTION: Railway CPU is the trainer (no GPU). Previous GPU comment retracted. P0-P2 still run locally at no cost; P3+ train on Railway CPU via the existing cpu_train / Dockerfile + entrypoint.sh path (AdamWCpu), exactly like the champion BPB=2.2111 run. See epic #1032 correction comment.
Decision metric locked: code BPB -> small pass@1 (MultiPL-E subset + t27-native executable eval). Generic-text BPB is diagnostic only.
Baseline frozen: champion BPB=2.2111 is the STANDARD AdamW arm (beta1=0.9, wd=0.04), NOT phi-canonical. Any phi claim must state this first.
CPU-bound reality recorded: trainer is cpu_train + AdamWCpu; no GPU.
Artefact: CODER_METRICS.md.
P1 (tokenizer + data spine) -- STARTED
Audited the real data contract: trainer is byte-level, VOCAB_SIZE=128, u8 targets, .bin token files (pipeline.rs).
Tokenizer decision (honest, CPU-bound): start byte-level, vocab 256 + 7 control/FIM sentinels (-> 263). The 32K-BPE SOTA default is GPU-era and is NOT used here; small 4K-8K BPE only if P4 CPU timing demands it. FIM slots reserved (<FIM_PRE/MID/SUF>), enabled at P2.
t27 parallel corpus extracted (build_t27_parallel_corpus.py -> t27_corpus/):
85 files, ~1.2 MB. Two compiler-verified translation signals:
numeric/formats_catalog emitted into 16 languages (single 16-way parallel module).
33 matched C<->Verilog module pairs across 10 categories (ar, base, compiler, fpga, isa, math, nn, numeric, queen, vsa).
HONEST SCALE: this is a seed + eval set, not pretraining volume (KBs). Value = aligned translation signal + executable eval, not token count.
Data spine tiers defined; proposed .bin contract + entrypoint env (no secrets): TRIOS_TRAIN_DATA=data/code_train.bin, TRIOS_VOCAB=263, TRIOS_HIDDEN=384, 3 parallel seed services {42,43,44} as the champion did.
Artefact: CODER_DATA_AND_TOKENIZER.md.
Next (P2): implement the binarizer tool (mirror trinity_3k_fineweb_train.rs), widen VOCAB_SIZE to 263, wire the real code next-token-prediction loss + code-BPB readout.
Activity
gHashTag commentedon Jun 1, 2026
Compute decision applied: P0-P2 run at $0 (local/sandbox). Start P0 now - no GPU needed. Railway = control plane + Postgres ledger + artefact storage; GPU = rented worker only from P3 onward. See epic #1032 compute-decision comment.
gHashTag commentedon Jun 1, 2026
CORRECTION: Railway CPU is the trainer (no GPU). Previous GPU comment retracted. P0-P2 still run locally at no cost; P3+ train on Railway CPU via the existing cpu_train / Dockerfile + entrypoint.sh path (AdamWCpu), exactly like the champion BPB=2.2111 run. See epic #1032 correction comment.
gHashTag commentedon Jun 1, 2026
P0 done + P1 started (2026-06-01) -- CPU/Railway
P0 (decision & honesty foundation) -- DONE
cpu_train+AdamWCpu; no GPU.CODER_METRICS.md.P1 (tokenizer + data spine) -- STARTED
VOCAB_SIZE=128,u8targets,.bintoken files (pipeline.rs).<FIM_PRE/MID/SUF>), enabled at P2.build_t27_parallel_corpus.py->t27_corpus/):numeric/formats_catalogemitted into 16 languages (single 16-way parallel module)..bincontract + entrypoint env (no secrets):TRIOS_TRAIN_DATA=data/code_train.bin,TRIOS_VOCAB=263,TRIOS_HIDDEN=384, 3 parallel seed services {42,43,44} as the champion did.CODER_DATA_AND_TOKENIZER.md.Next (P2): implement the binarizer tool (mirror
trinity_3k_fineweb_train.rs), widenVOCAB_SIZEto 263, wire the real code next-token-prediction loss + code-BPB readout.gHashTag commentedon Jun 16, 2026
P0 Complete — Closed
All sub-tasks delivered and verified:
Artefacts merged. Anchor: phi^2 + phi^-2 = 3.
fix(ci): Coq admitted check covers all .v files; add allow-list mecha…
docs: Wave Loop 14 — Loop 8 report, security hardening, CI fixes, res…