File size: 2,603 Bytes
c83e40d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 | {
"title": "GoLLeM-v5 64M — board-native honest status",
"updated_utc": "2026-09-24T17:00Z",
"source": "Hart recompute-verified (fabryka tiny_ml board-native harness); NOT lm-eval. Canonical: HF SlayerLab/gollem-v5-ckpts + live pods.",
"metric": "eff = mean(BLiMP%, ARC-Easy%, normWiki(BPB;1.86/500)) x sizeMult(64M=1.036)",
"board_context": {
"our_current_entry": {"rank": 16, "model": "GoLLeM-v5 32M", "eff": 75.51},
"pollock_128m": {"rank": 14, "eff": 75.66},
"leader": {"rank": 1, "model": "JugnuLM-110M", "eff": 80.20}
},
"v1_muon_64m_arcmix_board_native": {
"note": "Muon optimizer, arcmix corpus (9.42B). Board-native, recompute-verified.",
"trajectory": [
{"step": 40000, "tokens_b": 1.31, "blimp": 74.76, "arc_easy": 43.98, "bpb": 1.113, "eff": 74.64},
{"step": 80000, "tokens_b": 2.62, "blimp": 76.49, "arc_easy": 44.78, "bpb": 1.081, "eff": 75.65},
{"step": 160000, "tokens_b": 5.24, "blimp": 77.49, "arc_easy": 48.15, "bpb": 1.052, "eff": 77.29}
],
"status_now": "eff 77.29 at 40% training already > #16 (75.51) and #14 (75.66) -> ~#11-13 board-native.",
"projection_full_400k": {"eff_range": [78.0, 80.5], "arc_proj": [50, 52], "blimp_proj": [78.5, 79], "rank": "top-5 to top-10; #1 on optimistic edge", "method": "3-point convex log-extrapolation + WSD decay ARC-late; range not point (uncertain)."}
},
"v2_muon_64m_fwe_dominant": {
"note": "#1-shot candidate: FWE-78/Cosmo-22 (10.3B) + z-loss + logit-cap + WSD. IN PROGRESS.",
"trajectory": [
{"step": 40000, "tokens_b": 1.31, "blimp": 73.56, "arc_easy": 42.34, "bpb": 2.223, "eff": 73.49, "flag": "inconclusive-yellow: confounded vs v1 (data+recipe, not clean data-ablation); WSD stable-phase, ARC gain back-loaded to decay >320k."}
],
"status": "honest in-progress; NOT a data-win yet. Real test @160k/decay."
},
"v2b_clean_ablation": {
"note": "arcmix + v2-config (byte-identical to v2 except data). Isolates FWE data-lever. Launches when pod-1 frees (post v1-full).",
"status": "queued"
},
"path_to_1": {
"bottleneck": "ARC-Easy",
"need": "ARC ~54 at BLiMP~79/ppl~2.0 (BLiMP saturates ~79, ppl near floor -> #1 goes through ARC).",
"v1_projected_arc": [51, 52],
"gap": "~2-3 ARC",
"levers": ["v2 FWE-data (if +2-3 ARC over arcmix)", "v2b clean-test (does FWE raise ARC)", "distill synthetic-CoT (+1-3 ARC, next lever)"]
},
"discipline": "Ranges not points. Recompute-verified numbers only. Announce #1 only on board-native maintainer-reproducible numbers (avoid the old lm-eval +10-rank inflation)."
}
|