aphasia_adapter —— WAB 失语症评分 LoRA 适配器

基本信息

项	值
基座模型（base）	Qwen/Qwen3-32B-AWQ（AWQ 4-bit，Hugging Face 官方仓库）
LoRA rank / alpha	16 / 32（lora_target: all，q/k/v/o/gate/up/down_proj）
训练超参	lr=5e-05、epochs=5、grad_accum=4、cutoff_len=8192、seed=20260531
数据集	训练 179，用途=训练/训练FA0
适配器文件	adapter_config.json + adapter_model.safetensors（~512MB）
训练框架	LLaMA-Factory 0.9.3（QLoRA-on-AWQ，不设 quantization_bit）

质量指标（测试集 24 例）

与未微调 baseline 对比：

指标	baseline	本适配器
信息量 Spearman	0.738	0.933
信息量 QWK	0.756	0.880
流畅度 Spearman	0.659	0.864
流畅度 QWK	0.563	0.839
解析成功率	1.000	1.000

使用方式

以 PEFT 加载到 AWQ 基座上：

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen3-32B-AWQ"  # HF repo id：https://huggingface.co/Qwen/Qwen3-32B-AWQ
model = AutoModelForCausalLM.from_pretrained(
    base, trust_remote_code=True, torch_dtype="auto", device_map="auto"
)
model = PeftModel.from_pretrained(model, "artifacts/aphasia_adapter")

版权声明

本适配器（aphasia_adapter）是基于 Qwen/Qwen3-32B-AWQ （Apache-2.0 许可）的派生作品，继承其 Apache License 2.0 许可条款发布。使用、复制、修改或分发本适配器时，请遵守 Apache-2.0 的 相关义务（保留版权声明、声明实质性修改、不附带额外限制等）。训练数据来源及其使用授权不属本项目 可授予范围。

---

aphasia_adapter — WAB Aphasia Scoring LoRA Adapter

Overview

Item	Value
Base model	Qwen/Qwen3-32B-AWQ (AWQ 4-bit, official Hugging Face repo)
LoRA rank / alpha	16 / 32 (lora_target: all; q/k/v/o/gate/up/down_proj)
Training hyperparams	lr=5e-05, epochs=5, grad_accum=4, cutoff_len=8192, seed=20260531
Dataset	179 training cases, usage=训练/训练FA0
Adapter files	adapter_config.json + adapter_model.safetensors (~512MB)
Training framework	LLaMA-Factory 0.9.3 (QLoRA-on-AWQ, no quantization_bit)

Quality metrics (24-case test set)

Compared against the un-tuned baseline:

Metric	baseline	This adapter
Info Spearman	0.738	0.933
Info QWK	0.756	0.880
Fluency Spearman	0.659	0.864
Fluency QWK	0.563	0.839
Parse success rate	1.000	1.000

Usage

Load onto the AWQ base with PEFT:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen3-32B-AWQ"  # HF repo id: https://huggingface.co/Qwen/Qwen3-32B-AWQ
model = AutoModelForCausalLM.from_pretrained(
    base, trust_remote_code=True, torch_dtype="auto", device_map="auto"
)
model = PeftModel.from_pretrained(model, "artifacts/aphasia_adapter")

License

This adapter (aphasia_adapter) is a derivative work of Qwen/Qwen3-32B-AWQ, which is released under the Apache License 2.0. Accordingly, this adapter is distributed under the Apache License 2.0, inheriting the same terms. When using, copying, modifying, or redistributing this adapter, please comply with the Apache-2.0 obligations (preserve copyright notices, state material modifications, impose no additional restrictions, etc.). The dataset used for training is not covered by this license;