MiniSeek-V1-base

MiniSeek-M · V4.1-Flash 路线 · 预训练基座(非对话 / 非 SFT)

中文 | English

GitHub License Chat


🇨🇳 中文

超小开源 Decoder-only MoE 预训练基座,架构对齐 DeepSeek-V4.1-Flash 的小尺寸复现(--arch miniseek-m)。参数 186,933,744 ≈ 186.93M(Embedding / LM Head 不共享)。对话版见 MiniSeek-V1。

目标是对齐 DeepSeek(Dense → MoE → Flash 组件,base → chat),不是宣称已对标。

相对 Hub 上的 MiniSeek-M0.1-base(旧 1+32 top-2 MoELM),本版为全新结构,故记为 V1。

模型概览

Model ID wyf-yfw/MiniSeek-V1-base
系列 MiniSeek-M(MoE / Flash)
版本 1-base(仅预训练)
参数量 186,933,744(~186.93M)
权重文件 model.safetensors
结构 CED 编码4+解码4 · MoE 1 shared + 64 routed · top-6 · mHC · Engram · SWA/压缩索引
训练上下文 256 tokens
分词器 MiniMind(vocab_size=6400)+ ChatML chat_template
精度 BF16 训练 / FP32 model.safetensors
协议 MIT

结构要点

组件 取值
dim / n_layers 512 / 8
Attention 低秩 Q、共享 KV、滑窗+压缩索引(index topk=64)
MoE 1 shared + 64 routed,n_activated_experts=6,moe_inter_dim=176
路由 sqrt-softplus + bias,route_scale=1.5,clamp SwiGLU=10
mHC 4 流 · Sinkhorn 20
Engram layers 1, 3
vocab_size 6400

本检查点

项 说明
来源 task tasks/20261030-172542-ba2e/checkpoints/last.pt
Train step 10000
Train loss(ckpt) ≈ 3.7117
Valid loss @10k ≈ 3.05
数据 data/PT/pretrain_mix_12g_clean.jsonl
max_seq_len / batch 256 / 24 · pack
优化器 AdamW · LR 3e-4 → 1e-6 · warmup 200 · aux_coef=1e-4

代码:github.com/wyf-yfw/miniseek

快速开始(推荐 AutoModelForCausalLM)

已支持标准 Hugging Face AutoModelForCausalLM(无需依赖本地代码库,直接开箱即用):

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "wyf-yfw/MiniSeek-V1-base"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    trust_remote_code=True,
    dtype="auto",
    device_map="auto" if torch.cuda.is_available() else None,
)

inputs = tokenizer("春天来了,桃花", return_tensors="pt")
if torch.cuda.is_available():
    inputs = {k: v.cuda() for k, v in inputs.items()}

outputs = model.generate(
    **inputs,
    max_new_tokens=64,
    temperature=0.8,
    top_k=50,
    eos_token_id=tokenizer.eos_token_id,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

局限

  • ~187M、仅 10k PT step / 256 上下文,续写仍早期
  • 无指令 SFT(见 V1 chat)
  • 停止符 <|im_end|>

引用

@misc{miniseek-m1-base,
  title        = {MiniSeek-V1-base},
  author       = {wyf-yfw},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/wyf-yfw/MiniSeek-V1-base}},
  note         = {Ultra-small MoE / V4.1-Flash-style pretrained base}
}

🇬🇧 English

Ultra-small open decoder-only MoE pretrained base following a DeepSeek-V4.1-Flash–style stack (--arch miniseek-m). 186,933,744 ≈ 186.93M params (embedding not tied to LM head). Chat/SFT: MiniSeek-V1.

Goal: align toward DeepSeek. Not a claim of parity.

Succeeds Hub MiniSeek-M0.1-base (older 1+32 top-2 MoELM) as V1.

Summary

Parameters 186,933,744 (~186.93M)
MoE 1 shared + 64 routed, top-6, inter=176
Train context 256
Checkpoint task tasks/20261030-172542-ba2e/checkpoints/last.pt step 10000, train loss ≈ 3.7117
License MIT

See Chinese section for architecture table and load snippet.

Downloads last month
853
Safetensors
Model size
0.2B params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for wyf-yfw/MiniSeek-V1-base

Finetunes
1 model

Collection including wyf-yfw/MiniSeek-V1-base