Long-context sliding window: val_bpb=1.1764#61
Open
saml212 wants to merge 1 commit intoopenai:mainfrom
Open
Conversation
Novel finding: aggressive LR decay (WARMDOWN_ITERS=20000) reduces int8 quantization penalty from 0.014 to 0.005 BPB. Combined with FP16 tied embeddings and moderate NTK-RoPE extrapolation (eval@1408). Full warmdown sweep across 10 values and detailed analysis in README.
saml212
changed the title
saml212
changed the title
saml212
changed the title
saml212
changed the title
South-33
added a commit
to South-33/parameter-golf
that referenced
this pull request
Mar 19, 2026
- add a PR-audit research log entry covering the clean takeaways from pull requests openai#36 through openai#70 - promote long-context training plus matching long-context eval as a first-class clean branch based on PR openai#61 and PR openai#63 - refine mixed-precision export notes to emphasize using int6/int8 byte savings to fund wider MLP capacity, based on PR openai#65 - update the current snapshot and research thesis so future agents do not over-focus on exporter-only ideas after the broader PR sweep
South-33
added a commit
to South-33/parameter-golf
that referenced
this pull request
Mar 19, 2026
- fix the PR-audit notes to attribute the long-context branch to PR openai#65 rather than PR openai#61 - record PR openai#61 as schedule-side evidence about long warmdown reducing quantization damage - keep the ideas backlog aligned with the actual GitHub PR content before using it for next-step decisions
saml212
force-pushed
the
sam/warmdown-quantization
branch
from
50225c5 to
4a65f69
Compare
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
val_bpb = 1.1764
Baseline: 1.2244. Improvement: -0.048 BPB / -0.087 nats.
15.88MB artifact. 600s training + 80s sliding window eval on 8xH100 SXM.
Approach
Train at 2048 tokens with low LR (0.02), high Muon momentum (0.99), and tight gradient clipping (0.3). Evaluate with overlapping sliding windows (stride=512) — every scored token sees 1536+ tokens of context, matching training length exactly.
Novel Findings
Train length is irrelevant with sliding window. Training at 2048 vs 4096 gives identical BPB (1.1764 vs 1.1765) when using sliding window eval. The window provides long context at eval — the model only needs to learn local patterns. Training at 2048 gets more steps and is strictly better.
Gradient clipping has a narrow sweet spot for long sequences. clip=0.3 beats both 0.0 (no clip) and 0.1 (too tight). Full sweep in README.
Batch=786K is optimal (vs default 524K). Swept from 393K to 1M.
Configuration
Full details, sweep data, and reproduction in README.md.