Atomight V2.5 · 1.7B

Reasoning-first · Zero benchmark contamination · Trained on a free Colab T4

We are excited to announce and show you all, our most powerful and capable model in the current and newest Atomight family variant (V2.5).

Use this model when you want explicit chain-of-thought before the final answer — complex debugging, multi-step planning, agentic workflows, and math- or reasoning-heavy tasks.

Other details for this model soon. Wait for further information and details.

Quick Facts


Base model	Qwen/Qwen3-1.7B (Apache 2.0)
Training method	GRPO via LoRA/PEFT, merged to 16-bit
Trained on	Free-tier Google Colab T4 — no paid compute
Training data	~2,000 curated samples per domain, 6 premium open datasets
Contamination	Zero — no benchmark data used in training
License	CC-BY-4.0

Training Data

Curated (not scraped) from frontier 2025-era open datasets across STEM, science, math, and code:

Domain	Dataset	License
STEM	Logics-STEM-SFT-Dataset-Open-5.3M	Mixed (aggregated open sources)
Science	MegaScience	CC-BY / Academic Use
Math	OpenMathReasoning	CC-BY-4.0
Math	DeepMath-103K	CC-BY-4.0
Code	OpenCodeInstruct	CC-BY-4.0
Code	Nemotron-SFT-Competitive-Programming-v2	NVIDIA Open Model License

No benchmark data was used in training. None of the sources above overlap with MMLU, GSM8K, HumanEval, MBPP, HellaSwag, WinoGrande, ARC-Challenge, or TruthfulQA. Every score below was earned, not leaked.

Benchmark results

These are the results of the benchmarks for Atomight-V2.5-1.7B, evaluated with lm-evaluation-harness.

Benchmark	Metric	Score
MMLU	Accuracy	55.68
GSM8K	Accuracy	69.60
ARC-Challenge	Accuracy (normalized)	43.00
HellaSwag	Accuracy (normalized)	60.43
WinoGrande	Accuracy	61.09
TruthfulQA MC2	Accuracy	45.89
HumanEval	Pass@1	40.24
MBPP	Pass@1	42.80

_{All scores self-reported; no training data overlaps with these benchmarks — see Training Data section above.}

How it compares with other small language models (we recommend verifying it, as the other data from other models came from a third-party sources)

Scores below for other models are drawn from their respective model cards / technical reports, not re-run by us. Provided for context only — evaluation harnesses and prompt formats differ across labs, so treat this as directional rather than exact.

Model	Params	MMLU	GSM8K	HumanEval	MBPP	HellaSwag	WinoGrande
Atomight-V2.5-1.7B	1.7B	55.68	69.60	40.24	42.80	60.43	61.09
SmolLM2-1.7B (base)	1.7B	49.46	67.14	47.68	51.87	57.95	66.35
Qwen2.5-1.5B (base)	1.5B	63.03	66.57	35.37	58.37	66.60	66.20
Qwen2.5-1.5B-Instruct	1.5B	61.78	74.30	51.83	56.81	—	—
InfiR-1B-Instruct	1.0B	50.22	70.90	58.54	56.03	—	—
Llama-3.2-1B-Instruct	1.0B	46.27	47.90	39.63	49.03	—	—
TinyLlama-1.1B	1.1B	~27	~9	~9	~27	~59	~61

Where Atomight-V2.5-1.7B leads: best-in-class GSM8K among comparable 1–2B base models, and MMLU well above SmolLM2-1.7B — achieved with a curated, contamination-free ~2K-per-category dataset rather than large-scale pretraining.

Where it trails: HumanEval and MBPP lag behind instruction-tuned peers like InfiR-1B-Instruct and Qwen2.5-1.5B-Instruct — expected, since Atomight-V2.5-1.7B is a base reasoning model without dedicated code SFT. HellaSwag also trails models trained on broader web/commonsense data.

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "NovatasticRoScript/Atomight-V2.5-1.7B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

messages = [
    {"role": "system", "content": "You are a reasoning model. Think step-by-step inside <thinking> tags, then give your final answer inside <answer> tags."},
    {"role": "user", "content": "If a train travels 60 miles in 45 minutes, what is its speed in mph?"}
]

inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Main citation

This model is a fine-tuned derivative of Qwen3-1.7B (Apache 2.0) and is released under CC-BY-4.0 in accordance with the licenses of its training datasets. If you use this model, kindly credit NovatasticRoScript/Atomight-V2.5-1.7B and the base model/datasets listed above. Part of the Atomight family — small models, curated data, no shortcuts.

Other citacion/s

@misc{qwen3technicalreport,
      title={Qwen3 Technical Report}, 
      author={Qwen Team},
      year={2025},
      eprint={2505.09388},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.09388}, 
}

Training method — this model was trained using GRPO (Group Relative Policy Optimization), introduced in:

@article{deepseekmath2024,
  title={DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models},
  author={Shao, Zhihong and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Song, Junxiao and Zhang, Mingchuan and Li, Y.K. and Wu, Y. and Guo, Daya},
  journal={arXiv preprint arXiv:2402.03300},
  year={2024}
}