English / Japanese
Mistral-0.5B-init is a randomly initialized Mistral checkpoint that reuses the configuration and tokenizer of instruction-pretrain/InstructLM-500M. It is the starting point of the English 0.5B decoder-configuration ablation, not a usable model.
transformers>=4.51.0
accelerate>=1.6.0
sentencepiece>=0.2.0
flash-attn>=2.7.3
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "iamtatsuki05/Mistral-0.5B-init"
model_kwargs = {
"torch_dtype": torch.bfloat16,
"attn_implementation": "flash_attention_2",
"device_map": "auto",
}
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, **model_kwargs)
prompt = "Hachiware from Chiikawa is"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(
**inputs,
max_new_tokens=256,
temperature=0.8,
top_p=0.9,
do_sample=True,
)
print(tokenizer.decode(output[0], skip_special_tokens=True))
Initialization checkpoints for the English 0.5B ablation that varies the decoder configuration. Each uses the tokenizer of its source model.
| ID | Architecture | #Param. | #Param. w/o Emb. |
|---|---|---|---|
| iamtatsuki05/Mistral-0.5B-init (this model) |
Mistral | 518M | 469M |
| iamtatsuki05/Qwen2.5-0.5B-init | Qwen2.5 | 494M | 357M |
This model is distributed under the MIT License.
@article{MIREI
title={同一条件下における Encoder/Decoder アーキテクチャによる文埋め込みの性能分析},
author={岡田 龍樹 and 杉本 徹},
journal={言語処理学会第 32 回年次大会 (NLP2026)},
year={2026}
}