Evolve LoRA adapters

This private repository stores the 24 final LoRA adapter directories used in the Evolve layered SWE-QA experiments. It contains Qwen3.5-9B, Qwen3.5-35B-A3B, and Gemma-4-26B-A4B adapters for Direct SFT and v4 layered SFT at H2, H4, H5, and H6.

Layout

adapters/
├── qwen35-9b/
│   ├── direct/{h2,h4,h5,h6}/
│   └── v4/{h2,h4,h5,h6}/
├── qwen35-35b-a3b/
│   ├── direct/{h2,h4,h5,h6}/
│   └── v4/{h2,h4,h5,h6}/
└── gemma4-26b-a4b/
    ├── direct/{h2,h4,h5,h6}/
    └── v4/{h2,h4,h5,h6}/

Each level directory is the complete training final_adapter directory. In addition to the loadable top-level adapter_model.safetensors and adapter_config.json, retained intermediate checkpoint subdirectories are included for reproducibility.

Base model	Direct	v4	Total
Qwen3.5-9B	4 adapters, ~12 GB	4 adapters, ~12 GB	8 adapters, ~24 GB
Qwen3.5-35B-A3B	4 adapters, ~6.3 GB	4 adapters, ~6.3 GB	8 adapters, ~12.6 GB
Gemma-4-26B-A4B	4 adapters, ~11.2 GB	4 adapters, ~11.2 GB	8 adapters, ~22.4 GB

Loading an adapter

from peft import PeftModel
from transformers import AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-9B")
model = PeftModel.from_pretrained(
    base,
    "Yingxuan/evolve",
    subfolder="adapters/qwen35-9b/v4/h6",
)

The adapter configuration records PEFT 0.19.1, LoRA rank 64, alpha 128, and dropout 0.05. Use the matching Qwen3.5 or Gemma base model for each adapter family.

Training code, processed data, evaluation predictions, metrics, and complete evaluation trajectories are maintained at zoe-yyx/evolve.