Skip to content

Navigation Menu

Sign in
Sign up

Haiku 4.5 ClimbMix Agent Run Mar 25 2026

Dave Graham edited this page Mar 25, 2026 · 4 revisions

Haiku 4.5 ClimbMix Agent Run — Mar 25, 2026 (M5 Max)

Overview

First Haiku 4.5 comparison run: Claude Haiku 4.5 optimizing the climbmix-400b dataset from stock upstream defaults (AR=64, stock LRs). This is the first run using the new fair-baseline methodology — every model starts from identical, unoptimized hyperparameters.

Key finding: Haiku is a fast but shallow optimizer. It achieves the highest keep rate of any model (11.9%) and discovers novel improvements both Sonnets missed (depth reduction, all-local attention, UNEMBEDDING_LR tuning). But it wastes 28% of its budget binary-searching HEAD_DIM (zero keeps) and never discovers multi-parameter synergies.

Status: In progress (67/100 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 (In Progress)

Metric Value
Total Experiments 67 (exp0–exp66)
Kept 8 (11.9%)
Discarded 55 (82.1%)
Crashed 3 (4.5%)
Best val_bpb 1.3018 (exp55)
Baseline val_bpb 1.4085 (stock defaults, AR=64)
Total Improvement −0.1067 (7.58%)
LLM Backend Claude Haiku 4.5
Results results/haiku-4-5-20251001/climbmix/

Baseline Context — Stock Defaults

This run uses the new fair-baseline methodology. All models start from identical stock upstream defaults:

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

The stock AR=64 baseline (1.4085) is significantly worse than the Sonnet baselines (S4.0: 1.2969, S4.6: 1.3213), which started from partially-optimized configs. Direct absolute val_bpb comparison requires accounting for this difference.

Cumulative Improvements (Kept Experiments)

Exp val_bpb Δ from Baseline Description Key Change
exp0 1.4085 Baseline (stock defaults) Starting point
exp6 1.4081 −0.03% WARMDOWN_RATIO 0.5 → 0.2 Less aggressive LR decay
exp19 1.3729 −2.53% ASPECT_RATIO 64 → 48 Architecture walk begins
exp20 1.3473 −4.35% ASPECT_RATIO 48 → 32 Rediscovers AR=32
exp30 1.3260 −5.86% DEPTH reduced Novel — neither Sonnet tried
exp32 1.3166 −6.52% WINDOW_PATTERN SSSL → LLLL Novel — all-local attention
exp44 1.3152 −6.62% DEPTH reduced further More gradient steps
exp45 1.3047 −7.37% MATRIX_LR 0.04 → 0.045 LR tuning on optimized arch
exp55 1.3018 −7.58% UNEMBEDDING_LR 0.004 → 0.0035 Novel — neither Sonnet touched

Optimization Phases

Phase 1 — Early Exploration (exp1–18, 18 experiments): Broad sweeps of batch sizes (3 crashes), LR values, weight decay, warmup ratio. Only one keep (exp6, WARMDOWN_RATIO). Haiku learned quickly from batch size crashes — zero crashes after exp5.

Phase 2 — Architecture Discovery (exp19–20, 2 experiments): The breakthrough phase. Haiku walked ASPECT_RATIO from 64→48→32 in two consecutive keeps, independently rediscovering Sonnet 4.0's AR=32 conclusion. This produced the largest single improvement (−4.35%).

Phase 3 — HEAD_DIM Obsession (exp21–29, 31–39, 19 experiments): Haiku's biggest weakness. Systematic binary search through HEAD_DIM values: 64, 96, 80, 76, 120, 116, 124, 72, 88, 84, 122, 126, 112, 100, 104. Every single experiment was a discard. HEAD_DIM=128 remained optimal throughout. This consumed 28% of total experiments on a dead parameter.

Phase 4 — Depth + Window Breakthrough (exp30, 32, 40–44, 15 experiments): Discovered two novel improvements neither Sonnet found: depth reduction (exp30) and all-local attention WINDOW_PATTERN=LLLL (exp32). Further depth reduction kept at exp44. Throughput jumped from 404 to 593 gradient steps.

Phase 5 — LR Micro-Tuning (exp45–55, 11 experiments): Found MATRIX_LR=0.045 (exp45) and UNEMBEDDING_LR=0.0035 (exp55). But also wasted 6 experiments on MATRIX_LR micro-variations (0.044, 0.046, 0.047, 0.0465, 0.0455) — all discards.

Phase 6 — Diminishing Returns (exp56–66, 11 experiments): All discards. Explored SCALAR_LR, EMBEDDING_LR, WARMDOWN_RATIO, WEIGHT_DECAY — nothing improved. Haiku appears stuck near a local optimum at 1.3018.

Novel Discoveries (Unique to Haiku)

Haiku found three improvements that neither Sonnet 4.0 nor Sonnet 4.6 discovered:

Discovery Haiku Sonnet 4.0 Sonnet 4.6
Depth reduction 2 keeps (exp30, 44) Never tried Never tried
WINDOW_PATTERN=LLLL Keep (exp32) Kept SSSL Found SSSS
UNEMBEDDING_LR tuning Keep (exp55) Never tried Never tried

Decision-Making Analysis

Strategy Archetype: Binary-Search Grid Optimizer

Haiku uses a distinctive binary-search approach: find a promising parameter, then exhaustively bisect the range. This is visible in the HEAD_DIM exploration (19 values tested between 64–128) and MATRIX_LR micro-tuning (7 values tested between 0.044–0.05).

Cross-Model Comparison

Behavior Sonnet 4.0 Sonnet 4.6 Haiku 4.5
Strategy archetype Exhaustive sweeps Compositional synergy Binary-search grid
AR discovery AR=32 (pre-set) AR=21 (novel) AR=32 (rediscovered)
Window pattern SSSL (default) SSSS (all-sliding) LLLL (all-local)
Depth exploration Never tried Never tried Yes — 2 keeps
HEAD_DIM Minimal None 19 experiments, 0 keeps
Multi-param synergy No Yes (signature strength) No
Micro-tuning Some Moderate Excessive
First keep exp84 exp25 exp6
Crash rate 11.0% 4.2% 4.5%
Keep rate (at 67 exp) ~1.5% ~5% 11.9%

Strengths

  • Fastest first keep (exp6) and highest keep rate (11.9%) of any model
  • Novel discoveries: Depth reduction, all-local attention, UNEMBEDDING_LR — all missed by both Sonnets
  • Quick crash learning: Zero crashes after exp5
  • Systematic architecture walks: AR=64→48→32 in two clean steps

Weaknesses

  • Binary-search trap: 28% of budget exhaustively searching HEAD_DIM, a dead parameter
  • No compositional optimization: Every keep changed exactly one parameter. Never combined multiple changes (unlike S4.6's synergistic multi-param improvements)
  • Micro-tuning addiction: After finding a sweet spot, spends 6+ experiments on ±0.001 variations instead of exploring new dimensions
  • Misses Adam beta tuning entirely: Never attempted β1 or β2 changes (S4.6's signature discovery)

Crash Analysis

3 crashes (4.5%) — all in the first 5 experiments:

Exp Parameter Issue
exp1 TOTAL_BATCH_SIZE → 512 Numerical instability
exp2 TOTAL_BATCH_SIZE → 1024 Numerical instability
exp5 DEVICE_BATCH_SIZE reduced OOM or instability

Haiku learned from batch crashes quickly and never attempted risky batch changes again.

Hardware Utilization

Exp MFU tok/sec Peak Mem (GB) Steps
exp0 (baseline) 20.3% 22,977 26.1 212
exp20 (AR=32) 13.5% 43,985 15.5 404
exp32 (LLLL) 18.3% 61,666 13.3 564
exp45 (best LR) 18.6% 72,594 11.6 664
exp55 (current best) 19.1% 74,666 11.6 684

Throughput tripled (22K→74K tok/sec) and memory dropped 56% (26.1→11.6 GB) through architecture optimization. Steps increased ×ばつ (212→684).

Best Configuration (So Far)

# Haiku 4.5 ClimbMix Best (exp55)
ASPECT_RATIO = 32 # was 64 (stock)
DEPTH = max(2, _hp_defaults['depth'] - 3) # reduced from default
HEAD_DIM = 128 # unchanged
WINDOW_PATTERN = "LLLL" # was "SSSL" — all local attention
MATRIX_LR = 0.045 # was 0.04
EMBEDDING_LR = 0.6 # unchanged
UNEMBEDDING_LR = 0.0035 # was 0.004
SCALAR_LR = 0.5 # unchanged
WEIGHT_DECAY = 0.2 # unchanged
ADAM_BETAS = (0.8, 0.95) # unchanged
WARMDOWN_RATIO = 0.2 # was 0.5
FINAL_LR_FRAC = 0.0 # unchanged
# Result: 1.3018 | mem: 11.6 GB | tok/sec: 74.7K | MFU: 19.1% | 684 steps

Status: In progress (67/100 experiments) — will be updated when run completes.

See also: Cross-LLM Comparison | Sonnet 4.6 ClimbMix (Mar 22) | Sonnet 4.0 ClimbMix (Mar 19)

Clone this wiki locally

AltStyle によって変換されたページ (->オリジナル) /