mirror of
https://github.com/ooyinet/WeClone.git
synced 2026-09-01 16:14:38 +08:00
更新vllm依赖版本至0.8.2
This commit is contained in:
xming521
@@ -14,13 +14,11 @@
|
||||
## 核心功能✨
|
||||
- 💫 涵盖打造数字分身的全链路方案，包括聊天数据导出、预处理、模型训练、部署
|
||||
- 💬 使用微信聊天记录微调LLM，让大模型有"那味儿"
|
||||
- 🎙️ 使用微信语音消息➕0.5B大模型实现高质量声音克隆 👉[WeClone-audio](https://github.com/xming521/WeClone/tree/master/weclone-audio)
|
||||
- 🔗 绑定到微信、QQ、Telegram、企微、飞书机器人，实现自己的数字分身
|
||||
|
||||
## 特性与说明📋
|
||||
|
||||
> [!TIP]
|
||||
> 新特性：[WeClone-audio](https://github.com/xming521/WeClone/tree/master/weclone-audio) 模块，支持对微信语音进行克隆。
|
||||
|
||||
|
||||
> [!IMPORTANT]
|
||||
> 0.2.0版本进行了全面重构，数据集目录和脚本路径全部进行了修改，拉取新代码后，`csv`文件夹放在`dataset`下，并且需要重新安装依赖。
|
||||
|
||||
+1
-1
@@ -41,7 +41,7 @@ sparktts = [
|
||||
"torchaudio>=2.6.0",
|
||||
"tqdm>=4.66.5",
|
||||
]
|
||||
main = ["llamafactory>=0.9.2", "openai==1.76.0", "vllm==0.8.0"]
|
||||
main = ["llamafactory>=0.9.2", "openai==1.76.0", "vllm==0.8.2"]
|
||||
dev = ["pytest", "pyright", "ruff"]
|
||||
|
||||
[project.scripts]
|
||||
|
||||
@@ -0,0 +1,72 @@
|
||||
import pytest
|
||||
from weclone.data.clean.get_score import adjust_score_tiered
|
||||
|
||||
# 定义通用的参数
|
||||
THRESHOLDS = [0.6, 0.3] # 置信度阈值：>=0.6 高, >=0.3 中, <0.3 低
|
||||
DOWNGRADE_LEVELS = [0, 1, 2] # 对应降级幅度：高->0级, 中->1级, 低->2级
|
||||
|
||||
THRESHOLDS_FINE = [0.7, 0.5, 0.3]
|
||||
DOWNGRADE_LEVELS_FINE = [0, 1, 2, 3] # 对应 >=0.7, >=0.5, >=0.3, <0.3
|
||||
|
||||
test_cases = [
|
||||
# 案例 1: 高置信度
|
||||
(5, [0.05, 0.05, 0.1, 0.1, 0.7], THRESHOLDS, DOWNGRADE_LEVELS, 5, "高置信度"),
|
||||
# 案例 2: 中等置信度
|
||||
(4, [0.1, 0.15, 0.2, 0.45, 0.1], THRESHOLDS, DOWNGRADE_LEVELS, 3, "中等置信度"),
|
||||
# 案例 3: 低置信度
|
||||
(4, [0.15, 0.2, 0.25, 0.25, 0.15], THRESHOLDS, DOWNGRADE_LEVELS, 2, "低置信度"),
|
||||
# 案例 4: 低置信度，但原始分较低
|
||||
(2, [0.3, 0.2, 0.2, 0.15, 0.15], THRESHOLDS, DOWNGRADE_LEVELS, 1, "低置信度，原始分较低"),
|
||||
# 案例 5: 边界情况 - 刚好等于高阈值
|
||||
(3, [0.1, 0.1, 0.6, 0.1, 0.1], THRESHOLDS, DOWNGRADE_LEVELS, 3, "边界情况 - 等于高阈值"),
|
||||
# 案例 6: 边界情况 - 刚好等于中阈值
|
||||
(3, [0.2, 0.2, 0.3, 0.15, 0.15], THRESHOLDS, DOWNGRADE_LEVELS, 2, "边界情况 - 等于中阈值"),
|
||||
# 案例 7: 细分阈值 - 中高置信度
|
||||
(4, [0.1, 0.1, 0.2, 0.55, 0.05], THRESHOLDS_FINE, DOWNGRADE_LEVELS_FINE, 3, "细分阈值 - 中高置信度"),
|
||||
# 案例 8: 细分阈值 - 中低置信度
|
||||
(4, [0.15, 0.15, 0.2, 0.35, 0.15], THRESHOLDS_FINE, DOWNGRADE_LEVELS_FINE, 2, "细分阈值 - 中低置信度"),
|
||||
# 案例 9: 概率和异常 (预期行为是打印警告并继续计算)
|
||||
(3, [0.1, 0.1, 0.5, 0.1, 0.1], THRESHOLDS, DOWNGRADE_LEVELS, 3, "概率和异常"),
|
||||
]
|
||||
|
||||
@pytest.mark.parametrize("initial_score, probabilities, thresholds, downgrade_levels, expected_score, description", test_cases)
|
||||
def test_adjust_score_tiered(initial_score, probabilities, thresholds, downgrade_levels, expected_score, description):
|
||||
""" 测试 adjust_score_tiered 函数在各种情况下的表现 """
|
||||
print(f"测试案例: {description}")
|
||||
print(f" 输入: score={initial_score}, probs={probabilities}, thresholds={thresholds}, levels={downgrade_levels}")
|
||||
adjusted_score = adjust_score_tiered(initial_score, probabilities, thresholds, downgrade_levels)
|
||||
print(f" 输出: adjusted_score={adjusted_score}, 预期: {expected_score}")
|
||||
assert adjusted_score == expected_score
|
||||
|
||||
# 测试非法输入
|
||||
def test_adjust_score_invalid_input():
|
||||
""" 测试非法输入是否按预期引发 ValueError """
|
||||
# initial_score 无效
|
||||
with pytest.raises(ValueError, match="initial_score 必须在 1 到 5 之间"):
|
||||
adjust_score_tiered(0, [0.2]*5, THRESHOLDS, DOWNGRADE_LEVELS)
|
||||
with pytest.raises(ValueError, match="initial_score 必须在 1 到 5 之间"):
|
||||
adjust_score_tiered(6, [0.2]*5, THRESHOLDS, DOWNGRADE_LEVELS)
|
||||
|
||||
# probabilities 长度无效
|
||||
with pytest.raises(ValueError, match="probabilities 列表必须包含 5 个元素"):
|
||||
adjust_score_tiered(3, [0.2]*4, THRESHOLDS, DOWNGRADE_LEVELS)
|
||||
with pytest.raises(ValueError, match="probabilities 列表必须包含 5 个元素"):
|
||||
adjust_score_tiered(3, [0.1]*6, THRESHOLDS, DOWNGRADE_LEVELS) # 总和也不为1
|
||||
|
||||
# # probabilities 和不为 1 (现在是警告，不抛异常)
|
||||
# with pytest.raises(ValueError, match="probabilities 中元素的和必须接近 1.0"):
|
||||
# adjust_score_tiered(3, [0.1]*5, THRESHOLDS, DOWNGRADE_LEVELS)
|
||||
|
||||
# downgrade_levels 长度无效
|
||||
with pytest.raises(ValueError, match="downgrade_levels 的长度必须比 thresholds 的长度多 1"):
|
||||
adjust_score_tiered(3, [0.2]*5, THRESHOLDS, [0, 1])
|
||||
with pytest.raises(ValueError, match="downgrade_levels 的长度必须比 thresholds 的长度多 1"):
|
||||
adjust_score_tiered(3, [0.2]*5, THRESHOLDS, [0, 1, 2, 3])
|
||||
|
||||
# thresholds 不是降序
|
||||
with pytest.raises(ValueError, match="thresholds 列表必须是降序排列的"):
|
||||
adjust_score_tiered(3, [0.2]*5, [0.3, 0.6], DOWNGRADE_LEVELS)
|
||||
|
||||
# downgrade_levels 包含负数
|
||||
with pytest.raises(ValueError, match="downgrade_levels 中的降级幅度不能为负数"):
|
||||
adjust_score_tiered(3, [0.2]*5, THRESHOLDS, [0, -1, 2])
|
||||
@@ -15,7 +15,6 @@
|
||||
import json
|
||||
from typing import Optional
|
||||
|
||||
import fire
|
||||
from transformers import Seq2SeqTrainingArguments
|
||||
|
||||
from llamafactory.data import get_dataset, get_template_and_fix_tokenizer
|
||||
@@ -33,7 +32,7 @@ if is_vllm_available():
|
||||
|
||||
def vllm_infer(
|
||||
model_name_or_path: str,
|
||||
adapter_name_or_path: str = None,
|
||||
adapter_name_or_path: Optional[str] = None,
|
||||
dataset: str = "alpaca_en_demo",
|
||||
dataset_dir: str = "data",
|
||||
template: str = "default",
|
||||
@@ -158,5 +157,3 @@ def vllm_infer(
|
||||
print("*" * 70)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
fire.Fire(vllm_infer)
|
||||
@@ -0,0 +1,61 @@
|
||||
import math # 引入 math 模块以使用 floor 函数（虽然当前版本未使用，但保留以备平滑调整等扩展）
|
||||
|
||||
def adjust_score_tiered(initial_score: int, probabilities: list[float], thresholds: list[float], downgrade_levels: list[int]) -> int:
|
||||
"""
|
||||
根据大模型给出评分时的概率，对原始评分进行分级置信度调整。
|
||||
|
||||
Args:
|
||||
initial_score: 大模型给出的原始评分 (整数 1 到 5)。
|
||||
probabilities: 包含 5 个评分 (1 到 5) 概率的列表。
|
||||
例如 [P(1), P(2), P(3), P(4), P(5)]。
|
||||
thresholds: 一个降序排列的概率阈值列表，定义置信度区间边界。
|
||||
例如 [0.6, 0.3]。
|
||||
downgrade_levels: 与 thresholds 对应的降级幅度列表，长度比 thresholds 多 1。
|
||||
定义了每个置信度区间的降级数。例如 [0, 1, 2]。
|
||||
|
||||
Returns:
|
||||
经过置信度调整后的最终评分 (整数 1 到 5)。
|
||||
|
||||
Raises:
|
||||
ValueError: 如果输入参数不合法（例如概率列表长度不对，阈值未降序等）。
|
||||
"""
|
||||
# --- 输入校验 ---
|
||||
if not (1 <= initial_score <= 5):
|
||||
raise ValueError("initial_score 必须在 1 到 5 之间。")
|
||||
if len(probabilities) != 5:
|
||||
raise ValueError("probabilities 列表必须包含 5 个元素。")
|
||||
# 检查概率和是否接近 1 (允许小的浮点误差)
|
||||
if not math.isclose(sum(probabilities), 1.0, abs_tol=1e-6):
|
||||
print(f"警告: 概率之和 {sum(probabilities)} 不接近 1.0。请检查概率来源。") # 打印警告而非直接报错
|
||||
# raise ValueError("probabilities 中元素的和必须接近 1.0。")
|
||||
if len(downgrade_levels) != len(thresholds) + 1:
|
||||
raise ValueError("downgrade_levels 的长度必须比 thresholds 的长度多 1。")
|
||||
if any(thresholds[i] < thresholds[i+1] for i in range(len(thresholds)-1)):
|
||||
raise ValueError("thresholds 列表必须是降序排列的。")
|
||||
if any(level < 0 for level in downgrade_levels):
|
||||
raise ValueError("downgrade_levels 中的降级幅度不能为负数。")
|
||||
|
||||
# --- 算法核心 ---
|
||||
# 1. 获取选中分数的概率
|
||||
# 列表索引从0开始，所以评分 s 对应的索引是 s-1
|
||||
try:
|
||||
p_chosen = probabilities[initial_score - 1]
|
||||
except IndexError:
|
||||
# 这个错误理论上不应发生，因为 initial_score 已校验在 1-5 之间
|
||||
raise ValueError(f"无法从 probabilities 列表获取索引 {initial_score - 1} 的值。")
|
||||
|
||||
# 2. 确定降级幅度
|
||||
downgrade = downgrade_levels[-1] # 默认为最低置信度区间的降级幅度
|
||||
# 遍历阈值列表 (从高到低)
|
||||
for i in range(len(thresholds)):
|
||||
if p_chosen >= thresholds[i]:
|
||||
downgrade = downgrade_levels[i] # 找到对应的置信度区间
|
||||
break # 停止遍历
|
||||
|
||||
# 3. 计算调整后的评分
|
||||
preliminary_score = initial_score - downgrade
|
||||
adjusted_score = max(1, preliminary_score) # 确保分数不低于 1
|
||||
|
||||
# 4. 返回结果
|
||||
return adjusted_score
|
||||
|
||||
Reference in New Issue
Block a user