mikuhhn1239

qwen3-8b-narrative-parsing-lora

Deploy Dedicated

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Get help setting up a custom Dedicated Endpoints.

Talk with our engineer to get a quote for reserved GPU instances with discounts.

README

License: apache-2.0

任务

输入: 编号叙事单元 [1] "..." [2] "..." ...
输出: {"labels": [{"unit_id": "N", "type": "dialogue|narration|thought|action|scene_description"}]}
测试集: 39 条

五种类型

Table
类型	含义
`dialogue`	对话
`narration`	叙述
`thought`	心理
`action`	动作
`scene_description`	场景描写

示例

markdown
输入:
  [1] "你怎么来了？"
  [2] 她愣了一下。
  [3] 其实我也不知道自己为什么会来。

输出:
  {"labels": [
    {"unit_id": "1", "type": "dialogue"},
    {"unit_id": "2", "type": "action"},
    {"unit_id": "3", "type": "thought"}
  ]}

加载

python
from transformers import AutoModelForCausalLM
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(
    "mikuhhn1239/qwen3-8b-novel-base-sft",
    torch_dtype="auto", device_map="auto",
)
model = PeftModel.from_pretrained(
    base, "mikuhhn1239/qwen3-8b-narrative-parsing-lora"
)

训练

markdown
基座: Qwen3-8B-Novel-Base-SFT (Stage1 全参 SFT, 72K 小说续写数据)
方法: LoRA (r=64, α=128, dropout=0.05)
数据: 616 条 (577 train / 39 val / 39 test)
框架: transformers Trainer + PEFT
优化器: AdamW (adamw_torch_fused), cosine schedule, warmup=5%
epoch: 5 | LR: 1e-4 | batch: 1×16(accum) | bf16 | max_length: 4096

版本历史

Table
版本	数据量	epochs	LR	JSON解析	类型准确率	说明
零基座	—	—	—	0%	0%	Qwen3-8B 原始完全不会
+Stage1	—	—	—	0%	0%	读完 669 本也不会
v1	56	3	2e-4	57.1%	25.0%	端到端（切分+分类）
v2	310	3	2e-4	2.6%	63.6%	只分类，引号冲突
v3.1	310	5	1e-4	2.6%	63.6%	max_new_tokens=256 截断 JSON
v3.2 ⭐	577	5	1e-4	100%	69.5%	引号修复 + 扩标 + tokens→1024