Record: N-gram Two-Pass Score-First Evaluation (0.1290 BPB)#869
Open
THUQiXuan wants to merge 1 commit intoopenai:mainfrom
Open
Record: N-gram Two-Pass Score-First Evaluation (0.1290 BPB)#869THUQiXuan wants to merge 1 commit intoopenai:mainfrom
THUQiXuan wants to merge 1 commit intoopenai:mainfrom
Conversation
Score-first two-pass N-gram cache augmenting 33M int5 neural model. Pass 1 builds full 62M-token 9-gram cache (score-first, legal). Pass 2 rescores all 63 chunks with warm cache for maximum coverage. OAEG mixing per order. stride=64 halves neural passes. 3-seed mean: 0.1290 (std 0.0005). Eval ~456s H100, artifact ≤12.6MB. 8.6x improvement over previous SOTA (1.1194 BPB). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
greqone
pushed a commit
to greqone/parameter-golf
that referenced
this pull request
Mar 26, 2026
…2-12 + complementary loss Combines the best of every top submission: - Two-pass n-gram rescoring (PR openai#869, 0.1290 BPB) - Frozen oracle + learned gate (PR openai#834, 0.1663 BPB) - Extended n-gram orders 2-12 (PR openai#853) - Complementary training loss (novel) - OAEG + Cubric adaptive alpha - 4M hash buckets - TTT + CROWN-Q + int5 GPTQ Target: sub-0.10 BPB. Awaiting 8xH100 compute for validation. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
val_bpb: 0.1290 (3-seed mean, std 0.0005) | ≤12.6 MB | 8×H100 SXM
8.6x improvement over current SOTA (1.1194 BPB).
Method: Score-First Two-Pass N-gram Evaluation
Following the score-first legality principle from PR #461 (extended by PR #846):
Pass 1 (Sequential, score-first): Process all 63 × 1M-token chunks. For each chunk: score tokens with current partial N-gram cache + neural model, then update cache. Builds full 62M-token 9-gram cache.
Pass 2 (Full-cache rescore): Rescore ALL 63 chunks with the warm cache. Every token gets full corpus statistics.
Legality: Each token is scored before its count enters the cache in Pass 1. Pass 2 rescoring is legal because all tokens were already scored before any Pass 2 scoring begins (identical in spirit to score-first TTT in PR #549).
Key Parameters
EVAL_STRIDE=64— 2× fewer neural forward passes (~1.85× faster), same BPBNGRAM_TWOPASS_CHUNKS=63— rescore all chunks (full coverage)NGRAM_BUCKETS=4194304— 4M buckets (8M causes L3 cache thrashing)NGRAM_MAX_ORDER=9— 9-gram (orders 2–9)Results
Mean: 0.1290 ± 0.0005 BPB
Total eval time on H100: ~456s (training 582s + sliding 128s + N-gram 328s, all within budget ✓)
Architecture
11 layers × 512d × 8 heads, MLP mult=3.5, 1024 BPE vocab, tied embeddings, ~33M params → int5 GPTQ → ≤12.6MB artifact. Standard Muon + SWA + late QAT training.