10L + PPM Full-Rescore Order-12 N-gram (0.3461 BPB)#912
Closed
Bortlesboat wants to merge 6 commits intoopenai:mainfrom
Closed
10L + PPM Full-Rescore Order-12 N-gram (0.3461 BPB)#912Bortlesboat wants to merge 6 commits intoopenai:mainfrom
Bortlesboat wants to merge 6 commits intoopenai:mainfrom
Conversation
Explores stacking eval-time techniques (neural cache, LoRA TTT) and quantization-aware training on top of the openai#1 recipe. QAT has an export mismatch bug resulting in high quantization penalty — submitting as non-record to document the approach for iteration.
Non-record submission. 10 layers, d=512, GQA 8H/4KV, mixed int5/int6 quantization + zstd-22. BigramHash(4096, dim=128), SmearGate, SWA(0.4). Mean of 3 seeds: 1.1507 +/- 0.0006 BPB. All artifacts under 16MB.
10L d=512, GQA 8H/4KV, LeakyReLU(0.5)^2, Partial RoPE, LN Scale, XSA last 4, Value Residual, EMA(0.997). Mixed int5/int6 + zstd-22. Eval: multi-order hashed n-gram backoff (orders 2-7) with entropy- adaptive alpha. Mean of 3 seeds: 0.9123 +/- 0.0003 BPB.
Renamed to reflect actual technique (n-gram backoff + entropy alpha). Removed old 1.1507 BPB seed logs. Added explicit compliance/legality section per competition conventions.
Two-pass eval: pass 1 builds order 2-11 n-gram cache with order-adaptive entropy gating, pass 2 rescores cold-cache early windows with full cache. Mean of 3 seeds: 0.5863 +/- 0.0002 BPB. All artifacts under 16MB. Total eval: 331s on 8xH100.
PPM-style all-order blend (orders 2-12) with escape probabilities instead of hard backoff. Decoupled two-pass: pass 1 stores model_p, pass 2 rescores ALL tokens with leave-one-out against full cache. np.bincount for fast cache build. Mean of 3 seeds: 0.3461 +/- 0.0015.
Author
|
Replacing with a cleaner PR that only touches this submission's folder. |
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
1 participant
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.
Record submission
val_bpb: 0.3461 (mean of 3 seeds, std 0.0015)
What's novel
PPM-style all-order blend. Instead of hard backoff where only the highest matching order contributes, this blends ALL matching orders (2-12) using escape probabilities from PPM compression theory. Each order's contribution is weighted by its escape probability:
escape = beta / (ctx_count + beta). The neural model absorbs remaining mass. More principled than single-order interpolation.Leave-one-out self-exclusion. Each token's own (context, target) count is subtracted from the cache before scoring, eliminating self-inclusion bias in full-rescore.
Eval pipeline
Architecture
Compliance