minimind

MiniMind (meet447/minimind)

~64M MiniMind-3 dense checkpoint from the English-first fork meet447/minimind.

This Hub repo is the living weight store. New pretrain / SFT / later English runs overwrite main so you can always pull the latest.

Chat with the current main weights: meet447/minimind-chat.

Current weights (T4, 2026-09-02)

Architecture: 8 × 768, GQA 8Q/4KV, vocab 6400, SwiGLU, QK-Norm, RoPE 1e6, tied embeddings, no MoE.

This run: 1 epoch packed pretrain on pretrain_t2t_mini.jsonl (Tesla T4, fp16, batch 96) then 1 epoch SFT on sft_t2t_mini.jsonl (batch 16, lr 3e-5). Trainer extras: sequence packing, 3% warmup, AdamW param groups, torch.compile.

Honest quality note: the tokenizer and mini datasets are still the upstream Chinese-heavy ones. These weights validate the fork trainer. They are not an English-from-scratch model yet. English tokenizer + FineWeb-Edu / SmolTalk data come next (docs/ROADMAP.md).

Load (Transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "meet447/minimind"
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True, torch_dtype="auto")
messages = [{"role": "user", "content": "Why is the sky blue?"}]
inputs = tok.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt")
out = model.generate(inputs, max_new_tokens=128)
print(tok.decode(out[0], skip_special_tokens=True))

Load (this repo's trainer)

# download native .pth into ./out
huggingface-cli download meet447/minimind pytorch/full_sft_768.pth --local-dir .
# then
python eval_llm.py --weight full_sft --hidden_size 768 --num_hidden_layers 8

Update this repo

From a checkout that has out/full_sft_768.pth (and optionally out/pretrain_768.pth):

export HF_TOKEN=hf_...
python scripts/push_to_hub.py --repo-id meet447/minimind

License: Apache 2.0. Architecture and tokenizer follow jingyaogong/minimind.