t0-s1-14B is a fine-tuned language model developed at the Alan Turing Institute as part of the
t0 research initiative. It is a replication of the
S1 work —
Qwen2.5-14B-Instruct fine-tuned on the s1K dataset using supervised fine-tuning (SFT) via TRL. The s1K dataset is a curated set of 1,000 high-quality reasoning traces designed to elicit test-time scaling behaviour in smaller language models.
This model can be used for text generation and reasoning tasks. It is intended to explore test-time scaling behaviour in small language models, following the methodology of the S1 paper.
The model can be used as a reasoning-capable base for downstream tasks or plugged into larger pipelines. It is part of the broader t0 research initiative into lean yet capable LLMs.
The model inherits biases from its base model (Qwen2.5-14B-Instruct) and from the s1K training dataset. The s1K dataset is small and curated for reasoning; performance on other tasks may vary.
Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. Outputs should be reviewed appropriately for the intended use case.
1# Use a pipeline as a high-level helper
2from transformers import pipeline
3
4pipe = pipeline("text-generation", model="alan-turing-institute/t0-s1-14B")
5messages = [
6 {"role": "user", "content": "Who are you?"},
7]
8pipe(messages)
1# Load model directly
2from transformers import AutoTokenizer, AutoModelForCausalLM
3
4tokenizer = AutoTokenizer.from_pretrained("alan-turing-institute/t0-s1-14B")
5model = AutoModelForCausalLM.from_pretrained("alan-turing-institute/t0-s1-14B")
6messages = [
7 {"role": "user", "content": "Who are you?"},
8]
9inputs = tokenizer.apply_chat_template(
10 messages,
11 add_generation_prompt=True,
12 tokenize=True,
13 return_dict=True,
14 return_tensors="pt",
15).to(model.device)
16
17outputs = model.generate(**inputs, max_new_tokens=40)
18print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))
1pip install vllm
2vllm serve "alan-turing-institute/t0-s1-14B"
This model was fine-tuned on the
s1K dataset — a curated set of 1,000 reasoning traces used in the original S1 work. See the
S1 paper and the
t0-1 repository for more information on the training data and procedure.
Supervised fine-tuning (SFT) via
TRL, replicating the methodology described in the
S1 paper. See the
t0-1 repository for full details on hyperparameters and evaluation.
1@article{chan2025retrieval,
2 title={Retrieval-augmented reasoning with lean language models},
3 author={Chan, Ryan Sze-Yin and Nanni, Federico and Lazauskas, Tomas and Wood, Rosie and Yong, Penelope and Tarassenko, Lionel and Girolami, Mark and Geddes, James and Duncan, Andrew},
4 journal={arXiv preprint arXiv:2508.11386},
5 year={2025}
6}
7
8@article{muennighoff2025s1,
9 title={s1: Simple Test-Time Scaling},
10 author={Muennighoff, Niklas and Yang, Zitong and Shi, Weijia and Li, Xiang Lisa and Fei-Fei, Li and Hajishirzi, Hannaneh and Zettlemoyer, Luke and Liang, Percy and Candès, Emmanuel and Hashimoto, Tatsunori},
11 journal={arXiv preprint arXiv:2501.19393},
12 year={2025}
13}
