Maxwe11y/qwen3-ala-adapters

Qwen3 ALA LoRA Adapters

This repository contains LoRA adapters trained from Qwen/Qwen3-8B for controlled research on verb-centric counterfactual logic adaptation. Base-model weights and training data are not included.

Included Adapters

Directory	Configuration
output-qwen3-auto-g5-64-3-1e-05-3-token-25-uni	Standard SFT, n=3
output-qwen3-auto-g5-64-3-1e-05-1.0-0.1-3-token-25-uni	Contrastive-only, alpha=0.1, n=3
output-qwen3-sft-g5-0.05-64-3-1e-05-3-token-25-uni	Adaptive-only, beta=0.05, n=3
output-qwen3-contrastive-g5-0.1-0.1-64-3-1e-05-1.0-0.5-3-token-25-uni	Contrastive adaptation, beta=0.1, alpha=0.5, n=3

Loading

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "Qwen/Qwen3-8B"
adapter = "Maxwe11y/qwen3-ala-adapters/output-qwen3-contrastive-g5-0.1-0.1-64-3-1e-05-1.0-0.5-3-token-25-uni"
tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)

Use these adapters only for controlled research and evaluation.