MiniSeek
Collection
Goal: align toward DeepSeek. Starting from the simplest Dense models and advancing step by step (Dense → MoE, base → chat). One repo per checkpoint. • 6 items • Updated
超小开源 Decoder-only MoE 对话模型。架构参数 186,933,744 ≈ 186.93M(Embedding / LM Head 不共享)。预训练基座:MiniSeek-V1-base。
目标是对齐 DeepSeek(Dense → MoE → Flash 组件,base → chat),不是宣称已对标。
相对 Hub
MiniSeek-M0.1(旧 MoELM SFT),本版为全新结构,故为 V1。
| Model ID | wyf-yfw/MiniSeek-V1 |
| 基座 | wyf-yfw/MiniSeek-V1-base |
| 系列 | MiniSeek-M |
| 版本 | 1(chat / SFT) |
| 参数量 | 186,933,744(~186.93M) |
| 权重文件 | model.safetensors |
| 结构 | 同基座(CED + MoE 1+64 top-6 + mHC + Engram) |
| SFT 上下文 | 256 tokens |
| 分词器 | MiniMind(vocab_size=6400)+ ChatML |
| 协议 | MIT |
| 阶段 | 说明 |
|---|---|
| 基座 | MiniSeek-V1-base · tasks/20261030-172542-ba2e/checkpoints/last.pt last.pt |
| SFT 数据 | data/SFT/sft_miniseek_v5_short256.jsonl |
| 发布权重 | task tasks/20261030-233201-0c2a/checkpoints/best.pt best.pt |
| Train step(best) | 36428 |
| Train loss(best) | ≈ 1.4276 |
max_seq_len / batch |
256 / 24 · pack |
| 优化器 | AdamW · peak LR 5e-5 → 1e-6 · warmup 100 |
| aux | aux_coef=1e-4 |
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"
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,
)
model.infer_candidate_pool = True # 启用分层稀疏候选池推理
messages = [{"role": "user", "content": "中国的首都是哪里?只答城市名。"}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, 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=0.7,
top_k=50,
eos_token_id=tokenizer.eos_token_id,
)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
@misc{miniseek-m1,
title = {MiniSeek-V1},
author = {wyf-yfw},
year = {2026},
howpublished = {\url{https://huggingface.co/wyf-yfw/MiniSeek-V1}},
note = {Ultra-small MoE / V4.1-Flash-style SFT chat}
}
Ultra-small MoE chat / SFT checkpoint (186,933,744 ≈ 186.93M). Base: MiniSeek-V1-base. Released from task tasks/20261030-233201-0c2a/checkpoints/best.pt best.pt @ step 36428 (train loss ≈ 1.4276). MIT. See Chinese section for load snippet.
Base model
wyf-yfw/MiniSeek-V1-base