Skip to content

Navigation Menu

Sign in
Sign up

Sonnet 4.6 SlimPajama Agent Run Mar 25 2026

Dave Graham edited this page Mar 25, 2026 · 1 revision

Sonnet 4.6 SlimPajama Agent Run — Mar 25, 2026 (M5 Max)

Overview

Fourth cross-generation LLM comparison run: Claude Sonnet 4.6 optimizing SlimPajama on the same hardware previously used by Claude Sonnet 4.0. SlimPajama is the hardest dataset to compress (highest baseline val_bpb) — a 7-source deduplicated web mix that proved resistant to optimization by both models.

Key finding: SlimPajama is optimization-resistant. Sonnet 4.6 found only 2 keeps in 101 experiments (2.0%) — the lowest keep rate of any Sonnet 4.6 run. Combined with Sonnet 4.0's 3 keeps, this dataset has proven the hardest to optimize across both model generations. The defaults that carry over from Cosmopedia-v2 (AR=21, β1=0.65, β2=0.955) are already near-optimal for this multi-source data.

Hardware

Spec Value
Chip Apple M5 Max
Unified Memory 64 GB
GPU Cores 40
Backend MLX
Time Budget 5 minutes per experiment

Summary Statistics

Metric Value
Total Experiments 101 (exp0–exp100)
Kept 2 (2.0%)
Discarded 97 (96.0%)
Crashed 2 (2.0%)
Best val_bpb 1.5267 (exp91)
Baseline val_bpb 1.5312 (defaults)
Total Improvement −0.0045 (0.30%)
LLM Backend Claude Sonnet 4.6
Results results/sonnet-4-6/slimpajama/

Baseline Context

Metric Sonnet 4.0 Sonnet 4.6
Baseline val_bpb 1.5410 (pre-optimized) 1.5312 (defaults)
Best val_bpb 1.5259 1.5267
Improvement −1.0% −0.30%
Keeps 3 (3.0%) 2 (2.0%)
Crashes 0 (0.0%) 2 (2.0%)

Both models find SlimPajama near-impossible to improve. Sonnet 4.0 edges out on absolute val_bpb (1.5259 vs 1.5267), but started from a worse baseline (1.5410 vs 1.5312). The optimization landscape for this multi-source dataset is remarkably flat.

Cumulative Improvements (Kept Experiments)

Exp val_bpb Δ from Baseline Description Key Change
exp0 1.5312 Baseline (defaults) Starting point
exp63 1.5277 −0.23% Combined β1=0.55, FINAL_LR_FRAC=0.01, WARMDOWN_RATIO=0.70 Triple synergy
exp91 1.5267 −0.30% ADAM β2 0.955→0.945 Lower β2 smoothing

Optimization Phases

Phase 1 — Fruitless exploration (exp1–exp62): 62 experiments, zero keeps. The agent tried everything: AR changes (14, 16, 28), batch size reductions, LR adjustments, momentum changes, warmdown schedule variations. Nothing beat the baseline. This is the longest dry spell in any run across both models.

Phase 2 — Combination breakthrough (exp63): After 62 consecutive discards, the agent finally found improvement by combining three modest changes: β1 0.65→0.55, FINAL_LR_FRAC 0.02→0.01, WARMDOWN_RATIO 0.66→0.70. None worked alone — synergy was required.

Phase 3 — β2 fine-tuning (exp64–exp100): With exp63's config established, the agent exhaustively tested every remaining parameter. Only one more keep: β2 0.955→0.945 (exp91). 38 more experiments confirmed the configuration was near-optimal.

Best Configuration

# Sonnet 4.6 SlimPajama Best (exp91)
ASPECT_RATIO = 21 # unchanged from defaults
MATRIX_LR = 0.068 # unchanged
EMBEDDING_LR = 0.30 # was 0.50 (inherited default was different)
SCALAR_LR = 0.13 # unchanged
WEIGHT_DECAY = 0.1 # unchanged
WARMDOWN_RATIO = 0.70 # was 0.66
FINAL_LR_FRAC = 0.01 # was 0.02
ADAM_BETAS = (0.55, 0.945) # was (0.65, 0.955)
# Result: 1.5267 | mem: 14.5 GB | tok/sec: 50.6K | MFU: 15.6% | 464 steps

See also: Cross-LLM Comparison | Cross-Dataset Comparison | Sonnet 4.0 SlimPajama (Mar 20)

Clone this wiki locally

AltStyle によって変換されたページ (->オリジナル) /