Skip to content

Sonnet 4.6 Stock ClimbMix Agent Run Mar 26 2026

Dave Graham edited this page Mar 26, 2026 · 1 revision

Sonnet 4.6 ClimbMix Agent Run (Stock Baseline) — Mar 26, 2026 (M5 Max)

Overview

Sonnet 4.6 stock-baseline run: Claude Sonnet 4.6 optimizing the climbmix-400b dataset from stock upstream defaults (AR=64, stock LRs). Part of the fair-baseline methodology — every model starts from identical, unoptimized hyperparameters.

Key finding: Sonnet 4.6 is the most systematic architecture explorer. It walked ASPECT_RATIO from 64 down to 19 in six consecutive keeps, discovered a unique WINDOW_PATTERN=SLLL, and reversed its own MATRIX_LR decision (0.04→0.06→0.04→0.035) as the architecture evolved. But it missed depth reduction entirely, leaving the final win to Haiku.

Status: Complete (100/100 experiments)

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 100 (exp0–exp99)
Kept 15 (15.0%)
Discarded 81 (81.0%)
Crashed 4 (4.0%)
Best val_bpb 1.3093 (exp91)
Baseline val_bpb 1.3859 (stock defaults, AR=64)
Total Improvement −0.0766 (5.53%)
LLM Backend Claude Sonnet 4.6
Results results/sonnet-4-6/climbmix/

Baseline Context — Stock Defaults

Parameter Stock Default
ASPECT_RATIO 64
HEAD_DIM 128
WINDOW_PATTERN SSSL
MATRIX_LR 0.04
EMBEDDING_LR 0.6
UNEMBEDDING_LR 0.004
SCALAR_LR 0.5
WEIGHT_DECAY 0.2
ADAM_BETAS (0.8, 0.95)
WARMDOWN_RATIO 0.5
FINAL_LR_FRAC 0.0

Cumulative Improvements (Kept Experiments)

Exp val_bpb Δ from Baseline Description Key Change
exp0 1.3859 Baseline (stock defaults) Starting point
exp2 1.3842 −0.12% MATRIX_LR 0.04 → 0.06 Higher LR (later reversed)
exp27 1.3829 −0.22% ASPECT_RATIO 64 → 48 Architecture walk begins
exp28 1.3811 −0.35% ASPECT_RATIO 48 → 36 Narrowing continues
exp29 1.3345 −3.71% ASPECT_RATIO 36 → 28 Major throughput jump
exp30 1.3267 −4.27% ASPECT_RATIO 28 → 20 Deeper architecture
exp33 1.3251 −4.39% ASPECT_RATIO 20 → 18 Near-optimal AR
exp44 1.3248 −4.41% ASPECT_RATIO 18 → 19 Fine-tuned AR upward
exp46 1.3245 −4.43% MATRIX_LR 0.06 → 0.05 LR walk-back begins
exp47 1.3212 −4.67% MATRIX_LR 0.05 → 0.04 Reverted to stock!
exp57 1.3208 −4.70% SCALAR_LR 0.5 → 0.4 Lower scalar LR
exp58 1.3178 −4.91% SCALAR_LR 0.4 → 0.3 Further reduction
exp60 1.3173 −4.95% WARMDOWN_RATIO 0.5 → 0.55 Slightly longer warmdown
exp69 1.3151 −5.11% UNEMBEDDING_LR 0.004 → 0.006 Output layer tuning
exp90 1.3106 −5.43% WINDOW_PATTERN SSSL → SLLL Novel pattern discovery
exp91 1.3093 −5.53% MATRIX_LR 0.04 → 0.035 Combined with SLLL

Optimization Phases

Phase 1 — Initial LR Exploration (exp1–26, 26 experiments, 1 keep): Broad sweeps of all learning rates, betas, weight decay, warmup, warmdown. Only one keep: MATRIX_LR→0.06 (exp2). Three crashes (depth+1, AR=80, batch size). This phase established that stock LRs were already near-optimal — the real gains would come from architecture.

Phase 2 — Architecture Walk (exp27–34, 8 experiments, 6 keeps): The breakthrough phase. S4.6 systematically walked ASPECT_RATIO: 64→48→36→28→20→18, with every step a keep. Then tested AR=14 (discard — too narrow) and AR=16 (discard). Settled on AR=18, then fine-tuned to AR=19 at exp44. This is S4.6's signature strength: methodical, patient architecture exploration.

Phase 3 — LR Reversal (exp35–56, 22 experiments, 3 keeps): Having changed the architecture, S4.6 realized the original LR boost was wrong. Walked MATRIX_LR back: 0.06→0.05→0.04 (reverting to stock). Also discovered SCALAR_LR=0.3. This self-correcting behavior — boosting a param, then undoing it when context changed — is unique to S4.6.

Phase 4 — Fine-Tuning Plateau (exp57–89, 33 experiments, 2 keeps): Explored WARMDOWN_RATIO (0.55 kept), UNEMBEDDING_LR (0.006 kept), betas (all discards), weight decay (all discards), HEAD_DIM=64 (discard), depth+1 (discard), various window patterns. Long plateau with diminishing returns.

Phase 5 — Window Pattern Discovery (exp87–91, 5 experiments, 2 keeps): Late-game breakthrough. Tested SSLL (discard), SSGL (crash), SLSL (discard), then found SLLL (keep at exp90). Combined with MATRIX_LR=0.035 for the final best at exp91.

Phase 6 — Final Sweep (exp92–99, 8 experiments, 0 keeps): Attempted MATRIX_LR variations, SCALAR_LR, WARMDOWN_RATIO, EMBEDDING_LR, UNEMBEDDING_LR — all discards. Optimum appears stable.

Decision-Making Analysis

Strategy Archetype: Systematic Architecture Explorer with Self-Correction

S4.6's defining characteristic is patient, systematic architecture search. The AR walk (64→48→36→28→20→18→19) tested every major step and fine-tuned the optimum. More remarkably, S4.6 reversed its own MATRIX_LR decision when the architecture changed — the only model to self-correct a previously successful change.

The MATRIX_LR Journey

Exp MATRIX_LR AR val_bpb Context
exp0 0.04 64 1.3859 Stock default
exp2 0.06 64 1.3842 Boosted (keep)
exp46 0.05 19 1.3245 Walking back
exp47 0.04 19 1.3212 Reverted to stock!
exp91 0.035 19 1.3093 Below stock — architecture needed lower LR

S4.6 understood that a narrower architecture (AR=19) needed gentler weight updates. This adaptive LR strategy is absent in both S4.0 and Haiku.

AR Discovery Comparison

Model AR Walk Final AR Experiments Used
Sonnet 4.0 Never explored 64 (stock) 0
Sonnet 4.6 64→48→36→28→20→18→19 19 8 (6 keeps)
Haiku 4.5 64→48→32 32 2 (2 keeps)

S4.6 explored AR most thoroughly and found the narrowest optimal value. Its AR=19 is close to the AR=21 it found in pre-optimized runs, confirming this is a consistent S4.6 preference.

Crash Analysis

4 crashes (4.0%):

Exp Parameter Issue
exp3 DEPTH + 1 OOM or instability
exp16 ASPECT_RATIO 64 → 80 Model too wide
exp25 TOTAL_BATCH_SIZE reduced Numerical instability
exp88 WINDOW_PATTERN SSGL Invalid pattern crash

Crashes span the full run (exp3–88), unlike Haiku (all in first 5). S4.6 occasionally takes risky exploratory steps.

Hardware Utilization

Exp MFU tok/sec Peak Mem (GB) Steps
exp0 (baseline AR=64) 22.4% 25,291 26.1 233
exp29 (AR=28) 13.9% 45,272 15.5 414
exp33 (AR=18) 14.3% 46,651 15.5 427
exp47 (MATRIX_LR=0.04) 15.6% 50,616 15.5 465
exp91 (final best) 18.4% 51,615 15.5 474

Throughput doubled (25K→52K tok/sec) and memory dropped 41% (26.1→15.5 GB) through AR optimization. Steps increased 2.0× (233→474). Compare: Haiku achieved 76K tok/sec and 697 steps via depth reduction.

Best Configuration (Final)

# Sonnet 4.6 ClimbMix Best (exp91) — Stock Baseline
ASPECT_RATIO = 19        # was 64 (stock) — walked 64→48→36→28→20→18→19
HEAD_DIM = 128           # unchanged
WINDOW_PATTERN = "SLLL"  # was "SSSL" — novel pattern
MATRIX_LR = 0.035        # was 0.04 — went UP to 0.06, then back DOWN below stock
EMBEDDING_LR = 0.6       # unchanged
UNEMBEDDING_LR = 0.006   # was 0.004
SCALAR_LR = 0.3          # was 0.5
WEIGHT_DECAY = 0.2       # unchanged
ADAM_BETAS = (0.8, 0.95)  # unchanged (tried betas, no keeps)
WARMDOWN_RATIO = 0.55    # was 0.5
FINAL_LR_FRAC = 0.0      # unchanged
# Result: 1.3093 | mem: 15.5 GB | tok/sec: 51.6K | MFU: 18.4% | 474 steps

The Self-Correcting Optimizer

S4.6's run reveals a unique optimization personality: it adapts its strategy as context changes. The MATRIX_LR journey (stock → higher → reverted → below stock) shows that S4.6 doesn't just find good values — it re-evaluates past decisions when the landscape shifts.

This is fundamentally different from:

  • S4.0, which never changed architecture and micro-tuned the same parameters endlessly
  • Haiku, which binary-searched parameters in isolation until a late synergy emerged

S4.6's weakness was missing depth reduction entirely — a dimension that gave Haiku an extra 220+ gradient steps and the winning result.


Run complete (100/100 experiments). Sonnet 4.6 achieves val_bpb = 1.3093, 2nd place behind Haiku's 1.2953.

See also: Cross-LLM Comparison | Haiku 4.5 ClimbMix (Mar 25) | Sonnet 4.0 Stock ClimbMix (Mar 26)

Clone this wiki locally