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 cross-generation Haiku comparison run: Claude Haiku 4.5 optimizing the climbmix-400b dataset on the same hardware used by Claude Sonnet 4.0 and Sonnet 4.6. This adds a third model tier to the cross-LLM comparison matrix.

Status: In progress (22/101 experiments completed)

Early finding: Haiku's inherited baseline already beats both Sonnets' best. Starting from Sonnet 4.6's accumulated config (AR=21, β2=0.964), Haiku's unmodified baseline (1.2966) is already lower than Sonnet 4.0's best (1.2959, after 101 experiments) and Sonnet 4.6's best (1.2997, after 119 experiments). After just one keep (exp15), Haiku reaches 1.2924 — the new absolute best on ClimbMix.

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 22 (exp0–exp21)
Kept 1 (4.5%)
Discarded 19 (86.4%)
Crashed 2 (9.1%)
Best val_bpb 1.2924 (exp15)
Baseline val_bpb 1.2966 (inherited S4.6 config)
Total Improvement −0.0042 (0.32%)
LLM Backend Claude Haiku 4.5
Results results/haiku-4-5-20251001/climbmix/

⚠️ Baseline Context — Inherited Advantage

Unlike the Sonnet runs, Haiku started from Sonnet 4.6's accumulated best configuration (from the FineWeb-Edu-High run). This gives it a significant head start:

Metric Sonnet 4.0 Sonnet 4.6 Haiku 4.5
Baseline val_bpb 1.2969 (pre-optimized) 1.3213 (defaults) 1.2966 (S4.6 config)
Starting AR 32 32 21 (from S4.6)
Starting β2 0.95 0.95 0.964 (from S4.6)
Starting config Prior characterization Reset defaults S4.6 accumulated best

Haiku's baseline (1.2966) is already better than Sonnet 4.0's best-ever result (1.2959 after 101 experiments). Fair comparison requires accounting for this inherited advantage.

Inherited Starting Configuration

# From Sonnet 4.6's FineWeb-Edu-High accumulated best
ASPECT_RATIO = 21 # S4.6 discovery (was 32 for both Sonnets)
MATRIX_LR = 0.068 # S4.6 optimized
EMBEDDING_LR = 0.60 # S4.6 optimized
SCALAR_LR = 0.18 # S4.6 optimized
WEIGHT_DECAY = 0.08 # S4.6 optimized
WARMDOWN_RATIO = 0.75 # S4.6 optimized
FINAL_LR_FRAC = 0.05 # S4.6 optimized
ADAM_BETAS = (0.45, 0.964) # S4.6 β2 walk result
WINDOW_PATTERN = "SSSS" # S4.6 all-sliding

Cumulative Improvements (Kept Experiments)

Exp val_bpb Δ from Baseline Description Key Change
exp0 1.2966 Baseline (inherited S4.6 config) Starting point
exp15 1.2924 −0.32% SCALAR_LR 0.18 → 0.12 Lower scalar LR

Optimization Phases (So Far)

Phase 1 — Batch Size Exploration (exp1–3, 3 experiments): Haiku immediately tried to maximize gradient steps through batch size changes. TOTAL_BATCH_SIZE=256 crashed (exp1), DEVICE_BATCH_SIZE=8 degraded throughput (exp2), DEVICE_BATCH_SIZE=2 was slightly better but still worse than baseline (exp3). All batch experiments failed or degraded — the inherited config's batch settings are already near-optimal.

Phase 2 — LR and Schedule Sweeps (exp4–14, 11 experiments): Systematic exploration of MATRIX_LR, WARMDOWN_RATIO, WARMUP_RATIO, HEAD_DIM, DEVICE_BATCH_SIZE, and ASPECT_RATIO. All discards. Haiku tested AR=16 and AR=18 (both worse than inherited AR=21). HEAD_DIM=256 (exp11) came close (1.2994) but didn't beat baseline.

Phase 3 — Per-Parameter LR Tuning (exp15–21, 7 experiments): Found the only keep: SCALAR_LR 0.18→0.12 (exp15, val_bpb 1.2924). Then tried to extend the strategy to EMBEDDING_LR, UNEMBEDDING_LR, MATRIX_LR — all failed. SCALAR_LR 0.10 (exp21) overshot. Haiku correctly identified LR tuning as the right direction but struggled to find the next improvement.

Crash Analysis

2 crashes (9.1%) in 22 experiments:

Exp Parameter Issue
exp1 TOTAL_BATCH_SIZE → 256 Numerical instability from extreme batch reduction
exp6 DEVICE_BATCH_SIZE → 3 Odd batch size crash

Both crashes were in the first 6 experiments, from aggressive batch size changes. Haiku learned quickly and stopped attempting batch modifications after exp12.

Early Behavioral Observations

Behavior Sonnet 4.0 Sonnet 4.6 Haiku 4.5 (early)
First keep exp84 (84th) exp25 (25th) exp15 (15th)
Crash rate 11.0% 4.2% 9.1% (early)
Strategy Exhaustive sweeps Compositional LR micro-tuning
Batch experiments Many (11 crashes) Some (5 crashes) Few, stopped early
Unique strength Thoroughness Multi-param synergy Fast convergence

Haiku found its first keep faster than both Sonnets. However, it's struggling to find a second improvement — every experiment after exp15 has been a discard. The inherited config may be close enough to a local optimum that Haiku's LR-tuning strategy can't escape it.


Status: In progress — 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) /