Record: Fort Knox — Legal Packed Training Cache, Zero Val Adaptation (val_bpb 0.0638, 3-seed)#982

Closed

haikosys wants to merge 24 commits intoopenai:mainfrom

haikosys:fortknox-pr

Commits on Mar 26, 2026

Turbogrannie: TurboQuant + full-rescore n-gram (11L/576d/3.5x)

koltondrake
and
claude
committed
Fix submission structure to match leaderboard format

koltondrake
and
claude
committed
Fix torch.compile crash: @torch.compiler.disable on TurboQuant helpers

koltondrake
and
claude
committed
Bulletproof TurboQuant: extract entire QAT path out of compiled graph

koltondrake
and
claude
committed
fullgraph=False: allow graph breaks for @torch.compiler.disable

koltondrake
and
claude
committed
Safety fixes: weights_only=False + disable QAT before eval

koltondrake
and
claude
committed
13L default + suppress dynamo warnings + weights_only fix

koltondrake
and
claude
committed

Commits on Mar 27, 2026

Silence all dynamo recompile warnings

koltondrake
and
claude
committed
Rename submission folder 11L -> 13L to match actual config

koltondrake
and
claude
committed
Update README + submission.json for 13L with seed 1337 results

koltondrake
and
claude
committed
Turbocash: phrase cache + order-14 n-gram + 32M buckets + joint blend

koltondrake
and
claude
committed
Finalize turbogrannie: 3-seed results + submission package

koltondrake
and
claude
committed
README: TurboQuant claims vs reality commentary

koltondrake
and
claude
committed
Fiat: cache-first tiny model (6L/256d/3x, 4.2M params, FP16)

koltondrake
and
claude
committed
Fiat v2: interpolated multi-order n-gram + sequential phrase blend

koltondrake
and
claude
committed
Fiat v3: leave-one-out + greedy backoff + PR 913 alpha curves

koltondrake
and
claude
committed
CacheMoney: Full Tier 1+2+3 cache engine

koltondrake
and
claude
committed
Fix legacy single_pass compat: 3-tuple unpack + remove min_count kwarg

koltondrake
and
claude
committed
Fix cachemoney: pre-compute scores once, grid search only recomputes alpha

koltondrake
and
claude
committed
CacheMoney submission package: README + submission.json + seed 1337 log

koltondrake
and
claude
committed
CacheMoney final: 3-seed mean 0.0804 BPB (std 0.00003)

koltondrake
and
claude
committed
Fort Knox: zero val-adaptation baseline with packed training cache

koltondrake
and
claude
committed
Fort Knox: 0.0638 BPB, 3-seed, zero val adaptation, bulletproof legal

koltondrake
and
claude
committed
Clean PR branch: Fort Knox only

koltondrake
and
claude
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Record: Fort Knox — Legal Packed Training Cache, Zero Val Adaptation (val_bpb 0.0638, 3-seed)#982

Record: Fort Knox — Legal Packed Training Cache, Zero Val Adaptation (val_bpb 0.0638, 3-seed)#982
haikosys wants to merge 24 commits intoopenai:mainfrom
haikosys:fortknox-pr

Commits on Mar 26, 2026

Turbogrannie: TurboQuant + full-rescore n-gram (11L/576d/3.5x)

Fix submission structure to match leaderboard format

Fix torch.compile crash: @torch.compiler.disable on TurboQuant helpers

Bulletproof TurboQuant: extract entire QAT path out of compiled graph

fullgraph=False: allow graph breaks for @torch.compiler.disable

Safety fixes: weights_only=False + disable QAT before eval

13L default + suppress dynamo warnings + weights_only fix

Commits on Mar 27, 2026

Silence all dynamo recompile warnings

Rename submission folder 11L -> 13L to match actual config

Update README + submission.json for 13L with seed 1337 results

Turbocash: phrase cache + order-14 n-gram + 32M buckets + joint blend

Finalize turbogrannie: 3-seed results + submission package

README: TurboQuant claims vs reality commentary

Fiat: cache-first tiny model (6L/256d/3x, 4.2M params, FP16)

Fiat v2: interpolated multi-order n-gram + sequential phrase blend

Fiat v3: leave-one-out + greedy backoff + PR 913 alpha curves

CacheMoney: Full Tier 1+2+3 cache engine

Fix legacy single_pass compat: 3-tuple unpack + remove min_count kwarg

Fix cachemoney: pre-compute scores once, grid search only recomputes alpha

CacheMoney submission package: README + submission.json + seed 1337 log

CacheMoney final: 3-seed mean 0.0804 BPB (std 0.00003)

Fort Knox: zero val-adaptation baseline with packed training cache

Fort Knox: 0.0638 BPB, 3-seed, zero val adaptation, bulletproof legal

Clean PR branch: Fort Knox only