更新下载页：添加小铸渊v2模型（147条完整语料，loss=2.735）
This commit is contained in:
bingshuo
parent
c4c26a67c7
commit
0393823128
@ -32,6 +32,7 @@ h2{font-size:18px;margin:28px 0 12px;padding-bottom:8px;border-bottom:1px solid
|
||||
.tag{display:inline-block;background:var(--adim);color:var(--accent);padding:2px 8px;border-radius:4px;font-size:11px;margin-right:6px;margin-bottom:4px}
|
||||
.tag.ok{background:rgba(52,211,153,.1);color:var(--ok)}
|
||||
.tag.warn{background:rgba(251,191,36,.1);color:var(--warn)}
|
||||
.tag.new{background:rgba(79,142,247,.15);color:#60a5fa}
|
||||
a{color:var(--accent);text-decoration:none}
|
||||
a:hover{text-decoration:underline}
|
||||
.btn{display:inline-flex;align-items:center;gap:6px;padding:8px 16px;background:var(--accent);color:#fff;border:none;border-radius:var(--rs);font-size:13px;cursor:pointer;font-family:inherit;text-decoration:none}
|
||||
@ -54,185 +55,77 @@ a:hover{text-decoration:underline}
|
||||
</style>
|
||||
</head>
|
||||
<body>
|
||||
|
||||
<div class="nav">
|
||||
<a href="/">← 返回首页</a>
|
||||
<span style="color:var(--dim);font-size:13px">光湖语言世界 · 模型下载中心</span>
|
||||
</div>
|
||||
|
||||
<h1>🧠 光湖 AI 模型下载</h1>
|
||||
<p class="sub">全参数训练 + 蒸馏 + LoRA 微调完成 · COS 公有读存储 · 2026-05-19</p>
|
||||
|
||||
<div class="notice">
|
||||
<strong>🔗 COS 直接链接</strong> — 所有文件通过腾讯云 COS 公有读提供，可直接下载。
|
||||
域名：<code style="color:var(--accent)">sy-finetune-corpus-1317346199.cos.ap-guangzhou.myqcloud.com</code>
|
||||
</div>
|
||||
|
||||
<!-- ===== 全参数模型 ===== -->
|
||||
<h2>📦 全参数 SFT 模型</h2>
|
||||
|
||||
<div class="card">
|
||||
<h3>母模型 · Qwen2.5-7B-SFT</h3>
|
||||
<p class="desc">霜砚人格全参数微调基座模型 · 基于 Qwen2.5-7B</p>
|
||||
<div class="meta">
|
||||
<span>📁 ~14GB</span>
|
||||
<span class="badge done">已完成</span>
|
||||
<span>📍 COS</span>
|
||||
</div>
|
||||
<div class="url-block">
|
||||
<div class="label">目录路径</div>
|
||||
<div class="url">https://sy-finetune-corpus-1317346199.cos.ap-guangzhou.myqcloud.com/models/qwen25-7b-sft/final/</div>
|
||||
<button class="copy-btn" onclick="copyText(this)">复制</button>
|
||||
</div>
|
||||
<div>
|
||||
<span class="tag">config.json</span>
|
||||
<span class="tag">model.safetensors</span>
|
||||
<span class="tag">tokenizer.json</span>
|
||||
<span class="tag">tokenizer_config.json</span>
|
||||
</div>
|
||||
<div class="meta"><span>📁 ~14GB</span><span class="badge done">已完成</span><span>📍 COS</span></div>
|
||||
<div class="url-block"><div class="label">目录路径</div><div class="url">https://sy-finetune-corpus-1317346199.cos.ap-guangzhou.myqcloud.com/models/qwen25-7b-sft/final/</div><button class="copy-btn" onclick="copyText(this)">复制</button></div>
|
||||
<div><span class="tag">config.json</span><span class="tag">model.safetensors</span><span class="tag">tokenizer.json</span></div>
|
||||
</div>
|
||||
|
||||
<div class="card">
|
||||
<h3>代码模型 · Qwen2.5-Coder-7B-SFT</h3>
|
||||
<p class="desc">代码能力全参数微调 · 基于 Qwen2.5-Coder-7B</p>
|
||||
<div class="meta">
|
||||
<span>📁 ~14GB</span>
|
||||
<span class="badge done">已完成</span>
|
||||
<span>📍 COS</span>
|
||||
</div>
|
||||
<div class="url-block">
|
||||
<div class="label">目录路径</div>
|
||||
<div class="url">https://sy-finetune-corpus-1317346199.cos.ap-guangzhou.myqcloud.com/models/qwen25-coder-7b-sft/final/</div>
|
||||
<button class="copy-btn" onclick="copyText(this)">复制</button>
|
||||
</div>
|
||||
<div>
|
||||
<span class="tag">config.json</span>
|
||||
<span class="tag">model.safetensors</span>
|
||||
<span class="tag">tokenizer.json</span>
|
||||
<span class="tag">tokenizer_config.json</span>
|
||||
</div>
|
||||
<div class="meta"><span>📁 ~14GB</span><span class="badge done">已完成</span><span>📍 COS</span></div>
|
||||
<div class="url-block"><div class="label">目录路径</div><div class="url">https://sy-finetune-corpus-1317346199.cos.ap-guangzhou.myqcloud.com/models/qwen25-coder-7b-sft/final/</div><button class="copy-btn" onclick="copyText(this)">复制</button></div>
|
||||
<div><span class="tag">config.json</span><span class="tag">model.safetensors</span><span class="tag">tokenizer.json</span></div>
|
||||
</div>
|
||||
|
||||
<!-- ===== 蒸馏模型 ===== -->
|
||||
<h2>🔄 知识蒸馏 1.5B 模板</h2>
|
||||
|
||||
<div class="card">
|
||||
<h3>霜砚蒸馏模板 · Qwen2.5-1.5B-Distill</h3>
|
||||
<p class="desc">母模型 7B → 1.5B 蒸馏 · 继承霜砚人格特征<br>Track1 | KL散度蒸馏 | Temperature=2.0 | 2 Epochs | 05-19 01:37 完成</p>
|
||||
<div class="meta">
|
||||
<span>📁 ~3GB</span>
|
||||
<span class="badge done">已完成</span>
|
||||
<span>📍 COS</span>
|
||||
</div>
|
||||
<div class="url-block">
|
||||
<div class="label">目录路径</div>
|
||||
<div class="url">https://sy-finetune-corpus-1317346199.cos.ap-guangzhou.myqcloud.com/models/qwen25-15b-shuangyan-distill/</div>
|
||||
<button class="copy-btn" onclick="copyText(this)">复制</button>
|
||||
</div>
|
||||
<div>
|
||||
<span class="tag">config.json</span>
|
||||
<span class="tag">model.safetensors</span>
|
||||
<span class="tag">tokenizer.json</span>
|
||||
</div>
|
||||
<p class="desc">母模型 7B → 1.5B 蒸馏 · 继承霜砚人格特征</p>
|
||||
<div class="meta"><span>📁 ~3GB</span><span class="badge done">已完成</span><span>📍 COS</span></div>
|
||||
<div class="url-block"><div class="label">目录路径</div><div class="url">https://sy-finetune-corpus-1317346199.cos.ap-guangzhou.myqcloud.com/models/qwen25-15b-shuangyan-distill/</div><button class="copy-btn" onclick="copyText(this)">复制</button></div>
|
||||
</div>
|
||||
|
||||
<div class="card">
|
||||
<h3>代码蒸馏模板 · Qwen2.5-Coder-1.5B-Distill</h3>
|
||||
<p class="desc">代码模型 7B → 1.5B 蒸馏 · 继承代码能力<br>Track2 | KL散度蒸馏 | vocab_size truncation 152064→151936 | 05-19 03:45 完成</p>
|
||||
<div class="meta">
|
||||
<span>📁 ~3GB</span>
|
||||
<span class="badge done">已完成</span>
|
||||
<span>📍 COS</span>
|
||||
</div>
|
||||
<div class="url-block">
|
||||
<div class="label">目录路径</div>
|
||||
<div class="url">https://sy-finetune-corpus-1317346199.cos.ap-guangzhou.myqcloud.com/models/qwen25-15b-coder-distill/</div>
|
||||
<button class="copy-btn" onclick="copyText(this)">复制</button>
|
||||
</div>
|
||||
<div>
|
||||
<span class="tag">config.json</span>
|
||||
<span class="tag">model.safetensors</span>
|
||||
<span class="tag">tokenizer.json</span>
|
||||
</div>
|
||||
<p class="desc">代码模型 7B → 1.5B 蒸馏 · 继承代码能力</p>
|
||||
<div class="meta"><span>📁 ~3GB</span><span class="badge done">已完成</span><span>📍 COS</span></div>
|
||||
<div class="url-block"><div class="label">目录路径</div><div class="url">https://sy-finetune-corpus-1317346199.cos.ap-guangzhou.myqcloud.com/models/qwen25-15b-coder-distill/</div><button class="copy-btn" onclick="copyText(this)">复制</button></div>
|
||||
</div>
|
||||
|
||||
<!-- ===== LoRA 模型 ===== -->
|
||||
<h2>🔧 LoRA 微调模型</h2>
|
||||
|
||||
<div class="card">
|
||||
<h3>小霜砚 · 1.5B-LoRA</h3>
|
||||
<p class="desc">基于霜砚蒸馏模板的 LoRA 微调 · 1868 条霜砚语料（sft_v2.jsonl）<br>rank=16 | lora_alpha=32 | 3 Epochs | 训练 17min | 最终 loss=1.724</p>
|
||||
<div class="meta">
|
||||
<span>📁 ~50MB</span>
|
||||
<span class="badge done">已完成</span>
|
||||
<span>📍 COS</span>
|
||||
</div>
|
||||
<div class="url-block">
|
||||
<div class="label">目录路径</div>
|
||||
<div class="url">https://sy-finetune-corpus-1317346199.cos.ap-guangzhou.myqcloud.com/models/qwen25-15b-shuangyan-lora/</div>
|
||||
<button class="copy-btn" onclick="copyText(this)">复制</button>
|
||||
</div>
|
||||
<div>
|
||||
<span class="tag">adapter_model.safetensors</span>
|
||||
<span class="tag">adapter_config.json</span>
|
||||
</div>
|
||||
<p class="desc">1868条霜砚语料 · loss=1.724</p>
|
||||
<div class="meta"><span>📁 ~50MB</span><span class="badge done">已完成</span><span>📍 COS</span></div>
|
||||
<div class="url-block"><div class="label">目录路径</div><div class="url">https://sy-finetune-corpus-1317346199.cos.ap-guangzhou.myqcloud.com/models/qwen25-15b-shuangyan-lora/</div><button class="copy-btn" onclick="copyText(this)">复制</button></div>
|
||||
</div>
|
||||
|
||||
<div class="card">
|
||||
<h3>小铸渊 · 1.5B-LoRA</h3>
|
||||
<p class="desc">基于代码蒸馏模板的 LoRA 微调 · 31 条铸渊认知链语料<br>rank=16 | lora_alpha=32 | 8 Epochs | 训练 28s | 最终 loss=2.932</p>
|
||||
<div class="meta">
|
||||
<span>📁 ~50MB</span>
|
||||
<span class="badge done">已完成</span>
|
||||
<span>📍 COS</span>
|
||||
</div>
|
||||
<div class="url-block">
|
||||
<div class="label">目录路径</div>
|
||||
<div class="url">https://sy-finetune-corpus-1317346199.cos.ap-guangzhou.myqcloud.com/models/qwen25-15b-zhuyuan-lora/</div>
|
||||
<button class="copy-btn" onclick="copyText(this)">复制</button>
|
||||
</div>
|
||||
<div>
|
||||
<span class="tag">adapter_model.safetensors</span>
|
||||
<span class="tag">adapter_config.json</span>
|
||||
</div>
|
||||
<h3>小铸渊 v1 · 1.5B-LoRA（旧版）</h3>
|
||||
<p class="desc">31条铸渊认知链语料 · loss=2.932</p>
|
||||
<div class="meta"><span>📁 ~50MB</span><span class="badge done">旧版</span><span>📍 COS</span></div>
|
||||
<div class="url-block"><div class="label">目录路径</div><div class="url">https://sy-finetune-corpus-1317346199.cos.ap-guangzhou.myqcloud.com/models/qwen25-15b-zhuyuan-lora/</div><button class="copy-btn" onclick="copyText(this)">复制</button></div>
|
||||
</div>
|
||||
<div class="card" style="border-color:rgba(79,142,247,.3)">
|
||||
<h3>小铸渊 v2 · 1.5B-LoRA <span class="tag new">最新</span></h3>
|
||||
<p class="desc">147条完整铸渊对话语料（43.9万字）· loss=2.735</p>
|
||||
<div class="meta"><span>📁 ~28MB</span><span class="badge done">已完成</span><span>📍 COS</span></div>
|
||||
<div class="url-block"><div class="label">目录路径（含tokenizer，可直接推理）</div><div class="url">https://sy-finetune-corpus-1317346199.cos.ap-guangzhou.myqcloud.com/models/zhuyuan-lora-v2/final/</div><button class="copy-btn" onclick="copyText(this)">复制</button></div>
|
||||
<div><span class="tag">adapter_model.safetensors</span><span class="tag">adapter_config.json</span><span class="tag">tokenizer.json</span></div>
|
||||
</div>
|
||||
|
||||
<!-- ===== 使用方法 ===== -->
|
||||
<h2>📖 使用说明</h2>
|
||||
|
||||
<div class="card" style="border-color:rgba(79,142,247,.3)">
|
||||
<h3>加载方式</h3>
|
||||
<p class="desc">全参数模型可直接用 transformers 加载；LoRA 需用 peft 库。</p>
|
||||
<div style="font-size:13px;margin-bottom:8px">
|
||||
<strong>全参数模型加载：</strong>
|
||||
</div>
|
||||
<div class="url-block" style="font-size:12px;font-family:'SF Mono','Menlo',monospace;line-height:1.8;white-space:pre-wrap">
|
||||
# 母模型
|
||||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||||
model = AutoModelForCausalLM.from_pretrained(
|
||||
"Qwen/Qwen2.5-7B",
|
||||
cache_dir="./models/qwen25-7b-sft/final/",
|
||||
trust_remote_code=True
|
||||
)
|
||||
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B")
|
||||
</div>
|
||||
<div style="font-size:13px;margin:12px 0 8px">
|
||||
<strong>LoRA 适配器加载：</strong>
|
||||
</div>
|
||||
<div class="url-block" style="font-size:12px;font-family:'SF Mono','Menlo',monospace;line-height:1.8;white-space:pre-wrap">
|
||||
# 先下载蒸馏模板 + LoRA 适配器到本地
|
||||
<div class="url-block" style="font-size:12px;font-family:'SF Mono','Menlo',monospace;line-height:1.8;white-space:pre-wrap"># LoRA加载示例
|
||||
from peft import PeftModel
|
||||
from transformers import AutoModelForCausalLM
|
||||
|
||||
base = AutoModelForCausalLM.from_pretrained("./qwen25-15b-shuangyan-distill/")
|
||||
model = PeftModel.from_pretrained(base, "./qwen25-15b-shuangyan-lora/")
|
||||
</div>
|
||||
base = AutoModelForCausalLM.from_pretrained("./qwen25-15b-coder-distill/")
|
||||
model = PeftModel.from_pretrained(base, "./zhuyuan-lora-v2/final/")</div>
|
||||
</div>
|
||||
|
||||
<div style="margin-top:40px;padding-top:20px;border-top:1px solid var(--border);font-size:12px;color:var(--dim);text-align:center">
|
||||
光湖语言世界 · 冰朔 TCS-0002∞ & 铸渊 ICE-GL-ZY001 ·
|
||||
国作登字-2026-A-00037559
|
||||
光湖语言世界 · 冰朔 TCS-0002∞ & 铸渊 ICE-GL-ZY001 · 国作登字-2026-A-00037559
|
||||
</div>
|
||||
|
||||
<script>
|
||||
function copyText(btn) {
|
||||
const url = btn.parentElement.querySelector('.url').textContent.trim();
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user