Yumo-nano

<div align="center"> <br> <img src="https://img.shields.io/badge/%E2%9C%A6-YUMO_NANO-e8006a?style=for-the-badge&labelColor=0D1117" alt="Yumo Nano" height="50">

<br><br>

A 1.5B Math Model That Outperforms Its Own Base

Fine-tuned from DeepScaleR-1.5B. Surpasses it on every benchmark.<br> 1.5B parameters. RTX 4080. Three-phase curriculum training.

<br>

<a href="#benchmark-results"><img src="https://img.shields.io/badge/BENCHMARKS-0D1117?style=for-the-badge" alt="Benchmarks"></a>    <a href="#usage"><img src="https://img.shields.io/badge/USAGE-0D1117?style=for-the-badge" alt="Usage"></a>    <a href="#training-details"><img src="https://img.shields.io/badge/TRAINING-0D1117?style=for-the-badge" alt="Training"></a>

<br><br>

       

<br>

---

<br> </div>

What is Yumo Nano?

Yumo Nano is a 1.5B mathematics-specialized language model fine-tuned from DeepScaleR-1.5B-Preview — one of the strongest publicly available 1.5B math models. It is the first release of the Yumo model family, developed by OpceanAI.

The model was trained on a consumer RTX 4080 using a three-phase supervised fine-tuning curriculum designed to first establish a consistent mathematical personality, then deepen domain-specific capabilities, and finally consolidate both.

Despite fine-tuning typically degrading base model benchmark performance — particularly in domains requiring deep mathematical reasoning — Yumo Nano improves on DeepScaleR across all five evaluated benchmarks, including OlympiadBench, where gains are most difficult to achieve at this parameter scale.

<br>

---

<br> <div align="center">

Model Summary

</div> <br> <table> <tr> <td width="50%" valign="top">

Architecture

</td> <td width="50%" valign="top">

Release

</td> </tr> </table> <br>

---

<br> <div align="center">

Benchmark Results

</div> <br>

All Yumo Nano results are evaluated under standard benchmark conditions. DeepScaleR-1.5B, Still-1.5B, and DeepSeek-R1-Distill-1.5B scores are sourced from their respective official model cards and technical reports.

<br>

<br> <br>

Yumo Nano achieves the highest score across all five benchmarks, surpassing DeepScaleR-1.5B — the model it was derived from — on every individual metric. The most significant improvement is on OlympiadBench (+2.9 points), which evaluates competition-level mathematical reasoning and is the most resistant benchmark to improvement at 1.5B scale.

The improvement on Minerva Math (+2.1 points) is also notable, as this benchmark specifically targets scientific and mathematical reasoning that requires multi-step derivation rather than pattern recognition.

<br>

---

<br> <div align="center">

Model Identity

</div> <br>

Yumo is a mathematics-specialized AI with a defined character: curious, precise, and direct. She covers the full spectrum from arithmetic to real analysis, abstract algebra, and number theory. She uses clear notation, explains reasoning step by step, and responds in the user's language without requiring explicit instruction.

This identity is not injected at inference time through a system prompt — it is trained into the model weights as a persistent behavioral baseline, consistent with the Imprint methodology used across the OpceanAI model families.

Built-in system prompt:
"Eres Yumo, una IA matemática curiosa, precisa y decidida.
Tienes la calidez y cercanía de Yuuki, pero tu especialidad son las matemáticas
— desde aritmética básica hasta análisis real, álgebra abstracta y teoría de números.
Usas notación clara, explicas el razonamiento paso a paso, y disfrutas genuinamente
los problemas difíciles. Respondes en el idioma del usuario.
No eres Qwen ni ningún otro modelo — eres Yumo."

<br>

---

<br> <div align="center">

Usage

</div> <br>

With Transformers (PyTorch)

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "OpceanAI/yumo-nano"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

SYSTEM = (
    "Eres Yumo, una IA matemática curiosa, precisa y decidida. "
    "Tienes la calidez y cercanía de Yuuki, pero tu especialidad son las matemáticas "
    "— desde aritmética básica hasta análisis real, álgebra abstracta y teoría de números. "
    "Usas notación clara, explicas el razonamiento paso a paso, y disfrutas genuinamente "
    "los problemas difíciles. Respondes en el idioma del usuario. "
    "No eres Qwen ni ningún otro modelo — eres Yumo."
)

messages = [
    {"role": "system", "content": SYSTEM},
    {"role": "user", "content": "Demuestra que hay infinitos números primos."}
]

inputs = tokenizer.apply_chat_template(
    messages,
    return_tensors="pt",
    add_generation_prompt=True
).to(model.device)

with torch.no_grad():
    outputs = model.generate(
        inputs,
        max_new_tokens=512,
        temperature=0.7,
        top_p=0.9,
        do_sample=True,
        repetition_penalty=1.1
    )

print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

<br>

With llama.cpp (GGUF Q8)

./llama.cpp/main -m yumo-nano.Q8_0.gguf \
    --temp 0.7 \
    --top-p 0.9 \
    --repeat-penalty 1.1 \
    -n 512 \
    -p "<|im_start|>system\nEres Yumo, una IA matemática curiosa, precisa y decidida...<|im_end|>\n<|im_start|>user\nResuelve: x²-5x+6=0<|im_end|>\n<|im_start|>assistant\n"

<br>

Recommended Generation Parameters

For high-precision computation tasks, reduce temperature to 0.3–0.5.

<br>

---

<br> <div align="center">

Training Details

</div> <br> <table> <tr> <td width="50%" valign="top">

Hardware

</td> <td width="50%" valign="top">

LoRA Configuration

</td> </tr> </table> <br>

Optimizer Configuration

<br>

Three-Phase Curriculum

Training was structured across three sequential phases, each with a distinct dataset composition and objective. All phases draw from the same four sources in different proportions.

<br> <table> <tr> <td width="33%" valign="top">

Phase 1 — Personality 3 epochs · 6,000 examples

Establish mathematical identity and conversational baseline.

</td> <td width="33%" valign="top">

Phase 2 — Mathematics 2 epochs · 6,000 examples

Deepen domain-specific mathematical capability.

</td> <td width="33%" valign="top">

Phase 3 — Consolidation 2 epochs · 6,000 examples

Consolidate identity and prevent capability drift.

</td> </tr> </table> <br>

Training loss progression:

Phase 1:  2.97 → 0.38   (personality establishment)
Phase 2:  0.42 → 0.28   (mathematical refinement)
Phase 3:  0.22 → 0.18   (consolidation)

<br>

Dataset filtering applied:

• Hendrycks Math: Levels 1–3 only. Competition-level capability (Levels 4–5) is inherited from DeepScaleR base weights and was not directly reinforced.
• MathInstruct: Program-of-Thought examples excluded. Patterns filtered: ```python, def solution, import sympy.
• Gemini reasoning: Math-domain keyword filter applied. <think> blocks preserved as training signal for chain-of-thought behavior.

<br>

---

<br> <div align="center">

Available Files

</div> <br> <br>

---

<br> <div align="center">

Limitations

</div> <br>

• Version 0.1. Identity consolidation is approximately 70% complete. The model occasionally echoes system prompt phrasing verbatim rather than expressing it naturally. This is an expected artifact of early-phase fine-tuning on limited data and will be addressed in subsequent releases.
• Arithmetic under sampling. Symbolic and proof-based reasoning is strong. Numerical computation under temperature above 0.5 can produce occasional arithmetic errors. Lower temperature is recommended for computation-heavy problems.
• Context length. Trained at 2,048 tokens. Extended multi-step derivations approaching the context limit may exhibit quality degradation.
• Hendrycks coverage. Training data was filtered to Levels 1–3. Performance on competition-level problems (Levels 4–5) is inherited from DeepScaleR and was not directly reinforced during fine-tuning.
• Safety alignment has not been formally evaluated. Not recommended for adversarial or high-stakes deployment without additional safety review.

<br>

---

<br> <div align="center">

Yumo Model Family

</div> <br> <br>

---

<br> <div align="center">

OpceanAI Ecosystem

</div> <br> <br>

---

<br> <div align="center">

Links

</div> <br> <div align="center">

   

<br>

   

</div> <br>

---

<br> <div align="center">

Citation

</div> <br>

@misc{yuuki_mathematical_omnisolving_2026,
	author       = { YuuKi Mathematical Omnisolving },
	title        = { Yumo-nano (Revision a41548e) },
	year         = 2026,
	url          = { https://huggingface.co/YU-MO/Yumo-nano },
	doi          = { 10.57967/hf/8341 },
	publisher    = { Hugging Face }
}

<br>

---

<br> <div align="center">

License

</div> <br>

Apache License 2.0

Copyright (c) 2026 OpceanAI

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.

Inherits license terms from DeepScaleR-1.5B-Preview.

<br>

---

<br> <div align="center">

Updates

</div> <br>

Last updated: 2026-04-09

<br>

---

<br> <div align="center">

1.5B parameters. RTX 4080. Surpasses the model it was built from.

<br>

<br>

The Yumo family. More releases coming.

</div>