Skip to content

Record: EGGROLL v2 — val_bpb 1.1161 (3-seed mean, std 0.0001)#1156

Open
haikosys wants to merge 1 commit intoopenai:mainfrom
haikosys:eggroll-submission
Open

Record: EGGROLL v2 — val_bpb 1.1161 (3-seed mean, std 0.0001)#1156
haikosys wants to merge 1 commit intoopenai:mainfrom
haikosys:eggroll-submission

Conversation

@haikosys
Copy link
Copy Markdown

Record: EGGROLL v2 — val_bpb 1.1161 (3-seed mean, std 0.0001)

val_bpb: 1.1161 | val_loss: 1.884 nats | ~15.3 MB | 8×H100 SXM | Legal TTT

Built on PR #1130 by @Gusanidas (Kitchen Sink V2)
Foundation: PR #549 by @abaybektursun

3-seed validation, all artifacts under 16,000,000 bytes, all training under 600s.

Results

Seed val_loss (nats) val_bpb Artifact (bytes)
42 1.8848 1.1163 15,227,040
1337 1.8844 1.1160 15,427,072
2024 1.8844 1.1161 15,398,172
Mean 1.1161
Std 0.0001

Novel Contribution: EGGROLL (Antithetic Ternary Bin Search)

Post-GPTQ quantization refinement that directly optimizes INT6 bin assignments against BPB loss during eval budget (60s).

Algorithm: For each step, pick a quantized weight tensor, select 1024 random indices, test shifting bins +1 and -1 (antithetic pair), keep whichever improves loss. Strictly additive — cannot degrade
quality.

Properties:

  • Zeroth-order optimization (no gradients) on discrete quantized indices
  • Complementary to GPTQ: Hessian-based initial quantization + evolutionary refinement
  • 6-14 bin improvements per seed
  • Runs during eval budget, zero training impact

Also adds missing eval_val_sliding_ttt call to PR #1130's eval pipeline.

Timing

Phase Duration Budget
Training (8458 steps @ 69ms) 586s 600s
GPTQ calibration 14s (within training)
EGGROLL refinement 60s 600s eval
Sliding window eval 77s 600s eval
Score-first TTT ~470s 600s eval

Co-Authored-By: Claude Opus 4.6 (1M context) noreply@anthropic.com

val_bpb: 1.1161 | val_loss: 1.884 nats | ~15.3 MB | 8×H100 SXM | Legal TTT

Seeds: 42=1.1163, 1337=1.1160, 2024=1.1161 | Mean=1.1161, Std=0.0001

Novel: EGGROLL Antithetic Ternary Bin Search — post-GPTQ bin refinement
Also: adds missing TTT call to PR openai#1130 eval pipeline

Built on PR openai#1130 by @Gusanidas, PR openai#549 by @abaybektursun

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant