LinMix - Qwen3.5-like 4B 40 Domains mbpp Target - humaneval BPB: leaderboard
LinMix compares separately trained, publicly released language models built from scratch with different training-data mixtures. Each result belongs to its exact trained checkpoint; lower bits per byte is better. Bits per byte (lower is better).
Metric: Bits per byte (lower is better). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | LinMix qwen3.5-4B-40domains-vertex-linear-mbpp | 0.26 | |
| 2 | LinMix qwen3.5-4B-40domains-dirichlet-exponential-mbpp | 0.28 | |
| 3 | LinMix qwen3.5-4B-40domains-baseline | 0.28 | |
| 4 | LinMix qwen3.5-4B-40domains-dirichlet-lightgbm-mbpp | 0.29 |
Interactive version: theaggregate.ai/benchmark?slug=linmix-qwen3-5-like-4b-40-domains-mbpp-target-humaneval-bpb · How It Works · Data refreshed daily, snapshot 2026-10-29.