Text Generation
Transformers
Chinese
English
character-level
chinese
gru
mamba
Mixture of Experts
titans
Instructions to use masterofthestars/tianxuan with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use masterofthestars/tianxuan with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="masterofthestars/tianxuan")# pip install -U transformers accelerate # Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("masterofthestars/tianxuan", device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use masterofthestars/tianxuan with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "masterofthestars/tianxuan" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "masterofthestars/tianxuan", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker
docker model run hf.co/masterofthestars/tianxuan
- SGLang
How to use masterofthestars/tianxuan with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "masterofthestars/tianxuan" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "masterofthestars/tianxuan", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "masterofthestars/tianxuan" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "masterofthestars/tianxuan", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }' - Docker Model Runner
How to use masterofthestars/tianxuan with Docker Model Runner:
docker model run hf.co/masterofthestars/tianxuan
File size: 6,724 Bytes
db7bc06 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 | /**
* 语义反思层:在字符级模型输出基础上,增加"语言结构感"
* 通过 n-gram 约束 + 标点控制 + 语义连贯性检查,修复字符级模型的"乱码"问题
*/
import * as fs from "fs";
import * as path from "path";
/** 高频搭配矩阵:记录字符对 (bigram) 的共现频率 */
interface BigramMatrix {
matrix: Map<string, number>; // key: "char1|char2", value: count
total: number;
charFreq: Map<string, number>;
}
/** 反思规则配置 */
interface ReflectionConfig {
/** 最小句长(字符数),低于此值不强制结束 */
minSentenceLen: number;
/** 最大句长,超过强制结束 */
maxSentenceLen: number;
/** 最大连续重复字符数 */
maxRepeatCount: number;
/** bigram 概率阈值,低于此值时压制 */
bigramThreshold: number;
/** 标点概率提升权重 */
punctWeight: number;
/** 是否启用反思层 */
enabled: boolean;
}
const DEFAULT_CONFIG: ReflectionConfig = {
minSentenceLen: 5,
maxSentenceLen: 25,
maxRepeatCount: 2,
bigramThreshold: 0.001,
punctWeight: 5.0,
enabled: true,
};
/**
* 构建 n-gram 统计
*/
function buildBigramStats(text: string): BigramMatrix {
const matrix = new Map<string, number>();
const charFreq = new Map<string, number>();
let total = 0;
for (let i = 0; i < text.length - 1; i++) {
const a = text[i];
const b = text[i + 1];
// 统计 char-bigram
const key = `${a}|${b}`;
matrix.set(key, (matrix.get(key) ?? 0) + 1);
total++;
// 统计单字符频率
charFreq.set(a, (charFreq.get(a) ?? 0) + 1);
}
// 最后一个字符
if (text.length > 0) {
const last = text[text.length - 1];
charFreq.set(last, (charFreq.get(last) ?? 0) + 1);
}
return { matrix, total, charFreq };
}
/**
* 获取 bigram 条件概率 P(b|a)
*/
function bigramProb(matrix: BigramMatrix, a: string, b: string): number {
const count = matrix.matrix.get(`${a}|${b}`) ?? 0;
const freqA = matrix.charFreq.get(a) ?? 0;
if (freqA === 0) return 0;
return count / freqA;
}
/**
* 反思层:检查并修正生成的字符
* @param char 待生成的字符
* @param context 已生成的上下文(最近 maxSentenceLen 字符)
* @param config 配置
* @param bigramStats n-gram 统计
*/
export function reflect(
char: string,
context: string,
config: ReflectionConfig = DEFAULT_CONFIG,
bigramStats?: BigramMatrix,
): {
char: string;
action: 'accept' | 'reject' | 'force_punct' | 'force_end';
reason: string;
} {
if (!config.enabled) {
return { char, action: 'accept', reason: 'reflection disabled' };
}
// 规则1: 连续重复检测
if (context.length >= config.maxRepeatCount) {
const suffix = context.slice(-config.maxRepeatCount);
if (suffix.split('').every(c => c === char)) {
// 强制插入句号,打破循环
return { char: '。', action: 'force_end', reason: `repeat break: ${char.repeat(config.maxRepeatCount + 1)}` };
}
}
// 规则2: 句长控制
const currentSentenceLen = getCurrentSentenceLen(context);
if (currentSentenceLen >= config.maxSentenceLen) {
return { char: '。', action: 'force_end', reason: `sentence too long: ${currentSentenceLen}` };
}
// 规则3: bigram 概率检测
if (bigramStats && context.length > 0) {
const lastChar = context[context.length - 1];
const prob = bigramProb(bigramStats, lastChar, char);
if (prob < config.bigramThreshold && char !== '。' && char !== ',' && char !== '!' && char !== '?') {
// 低概率组合,但有概率接受(模拟语言的不确定性)
if (Math.random() < 0.3) {
return { char, action: 'accept', reason: `low prob bigram: ${lastChar}${char}(${prob.toFixed(4)})` };
}
// 否则尝试替换为更高概率的字符
const forced = findBetterChar(context, bigramStats, config);
if (forced) {
return { char: forced, action: 'reject', reason: `low prob replaced: ${lastChar}${char}(${prob.toFixed(4)}) -> ${forced}` };
}
}
}
// 规则4: 标点概率提升
if (char === '。' || char === ',' || char === '!' || char === '?') {
return { char, action: 'accept', reason: 'punctuation' };
}
return { char, action: 'accept', reason: 'normal' };
}
/**
* 获取当前句子的长度(从最后一个标点之后开始计算)
*/
function getCurrentSentenceLen(context: string): number {
let len = 0;
for (let i = context.length - 1; i >= 0; i--) {
if (['。', '!', '?'].includes(context[i])) break;
len++;
}
return len;
}
/**
* 寻找更好的替代字符(基于 bigram 概率)
*/
function findBetterChar(context: string, stats: BigramMatrix, config: ReflectionConfig): string | null {
const lastChar = context[context.length - 1];
if (!lastChar) return null;
// 收集所有可能的下一个字符及其概率
const candidates: Array<{ char: string; prob: number }> = [];
const matrixEntries = Array.from(stats.matrix.entries());
for (const [key, count] of matrixEntries) {
const [a, b] = key.split('|');
if (a === lastChar) {
const freqA = stats.charFreq.get(a) ?? 0;
if (freqA > 0) {
candidates.push({ char: b, prob: count / freqA });
}
}
}
// 按概率排序
candidates.sort((a, b) => b.prob - a.prob);
// 选择概率最高的几个候选
const topCandidates = candidates.slice(0, 5);
if (topCandidates.length === 0) return null;
// 过滤掉标点,优先选择汉字
const hanziCandidates = topCandidates.filter(c => /^[\u4e00-\u9fa5]$/.test(c.char));
if (hanziCandidates.length > 0) {
return hanziCandidates[0].char;
}
return topCandidates[0].char;
}
/**
* 从语料构建 bigram 统计并保存
*/
export function buildAndSaveBigrams(corpusPath: string, outputPath: string): void {
const text = fs.readFileSync(corpusPath, 'utf-8');
const stats = buildBigramStats(text);
// 转换为可序列化的格式
const data = {
matrix: Array.from(stats.matrix.entries()),
total: stats.total,
charFreq: Array.from(stats.charFreq.entries()),
};
fs.writeFileSync(outputPath, JSON.stringify(data, null, 2));
console.log(`Bigram stats saved: ${stats.total} bigrams, ${stats.charFreq.size} unique chars`);
}
/**
* 加载 bigram 统计
*/
export function loadBigrams(path: string): BigramMatrix | null {
try {
const data = JSON.parse(fs.readFileSync(path, 'utf-8'));
return {
matrix: new Map(data.matrix),
total: data.total,
charFreq: new Map(data.charFreq),
};
} catch (e) {
console.warn(`Failed to load bigrams from ${path}: ${e}`);
return null;
}
}
|