Downloads · 30 days
16
3% of all-time downloads
kopriaLM
Ryosei-Akatsuka/kopriaLM
kopriaLM is a text generation model from Ryosei-Akatsuka. Use it when you need the model to write or continue text. The card lists the license as mit.
99,360 パラメータの極小言語モデル（再帰 Transformer + TT-Embedding + LoRA）。
All-time downloads
578
Public
Parameters
149K
597 KB on disk
Likes
1
Public
Repo makeup
Click a slice to open those files.
.safetensors597 KB · 74%
At a glance
- Task
- Text Generation
- License
- mit
- Model type
- kopria
- Access
- Public
- Created
- Aug 8, 2026
- Updated
- Aug 9, 2026
- SHA
- 82d2c48d
Try a prompt
Model card
From the Hugging Face model README
kopriaLM
99,360 パラメータの極小言語モデル（再帰 Transformer + TT-Embedding + LoRA）。
- 語彙: 3,000（ByteLevel BPE）
- コンテキスト長: 512
- 事前学習: BabyLM-2026-Strict-Small + WikiText-2
- SFT: databricks-dolly-15k（3 epoch）
ベンチマーク結果
| Task | Shots | Metric | Score |
|---|---|---|---|
| HellaSwag | 0 | acc_norm | 25.39% |
| LAMBADA OpenAI | 0 | acc | 0.10% |
| PIQA | 0 | acc_norm | 49.78% |
| WinoGrande | 0 | acc | 49.80% |
| ARC-Easy | 0 | acc_norm | 24.83% |
| ARC-Challenge | 0 | acc_norm | 26.11% |
| BoolQ | 0 | acc | 37.83% |
| MMLU（57科目平均） | 0 | acc | 23.02% |
| 全タスク平均 | - | - | 29.61% |
使い方
trust_remote_code=True が必要です。KVキャッシュ非対応のため use_cache=False で生成してください。
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
"Ryosei-Akatsuka/kopriaLM", trust_remote_code=True)
tok = AutoTokenizer.from_pretrained("Ryosei-Akatsuka/kopriaLM")
ids = tok("Q: What is 2+2?\nA: ", return_tensors="pt")
gen = model.generate(**ids, max_new_tokens=30, use_cache=False)
print(tok.decode(gen[0], skip_special_tokens=True))
ライセンス
MIT
- Task
- Text Generation
- Type
- kopria
- License
- mit
- Languages
- en
- Created
- Aug 8, 2026
- Updated
- Aug 9, 2026