cloudyu's picture
Add milder edit (alpha=0.05)
d979661 verified
|
Raw History Blame Contribute Delete
1.16 kB

hy4-flash-concise / alpha-0.05 (milder edit)

This is the same rank-one, gradient-free edit as the main version, applied at half strength (α = 0.05). Only model-wsp-edit.safetensors differs from the repository root. Every other file is identical.

It saves fewer tokens than α = 0.1, and accuracy on reasoning and code stays within noise of the base model.

Benchmark Base α=0.05 Δ accuracy [95% CI] Δ thinking tokens
GPQA Diamond (198 × 2) 91.2 89.6 −1.5 [−3.8, +0.5] −20.6%
LiveCodeBench v6 (175) 84.6 82.3 −2.3 [−8.0, +3.4] −25.0%
IFBench (300, loose) 76.7 70.3 −6.3 [−11.0, −2.0] −36.6%
13 held-out tasks × 3 39/39 37/39 −5.1 −29.1%

Download only this folder and serve it as a local path:

hf download autotrust/hy4-flash-concise --include "alpha-0.05/*" --local-dir hy4-flash-concise
vllm serve hy4-flash-concise/alpha-0.05 --served-model-name hy4-flash --reasoning-parser hy_v4 ...

For llama.cpp, use the root gguf-lora/hy4-flash-concise-lora.gguf with scale 0.05: --lora-scaled hy4-flash-concise-lora.gguf:0.05.