Model Card for t0-1.1-k5-32B
t0-1.1-k5-32B is a fine-tuned language model developed at the Alan Turing Institute as part of the t0 research initiative, which focuses on lean yet highly capable LLMs for Retrieval-Augmented Reasoning (RAR). This model is fine-tuned from Qwen2.5-32B-Instruct and optimised for use in a RAG pipeline applied to a domain-specific body of knowledge (demonstrated on NHS A-to-Z condition webpages). The k5 in the model name denotes top-k=5 retrieval during training/evaluation.
Model Details
Model Description
- Developed by: t0 team at the Alan Turing Institute
- Authors: Ryan Sze-Yin Chan, Federico Nanni, Tomas Lazauskas, Rosie Wood, Penelope Yong, Lionel Tarassenko, Mark Girolami, James Geddes, Andrew Duncan
- Model type: Text Generation (causal language model)
- Language(s) (NLP): English
- License: Apache 2.0
- Finetuned from model: Qwen2.5-32B-Instruct
Model Sources
Uses
Direct Use
This model can be used directly for text generation and conversational tasks, particularly question answering in the context of health-related information. It is designed to work as the generator component of a Retrieval-Augmented Generation (RAG) pipeline.
Downstream Use
The model is intended to be plugged into a RAG system (as described in the t0-1 repository) where it reasons over retrieved documents from a domain-specific knowledge base. The reference application uses NHS A-to-Z condition webpages as the knowledge source.
Out-of-Scope Use
This model is not intended for:
- Medical diagnosis or clinical decision-making without appropriate oversight
- Use cases outside of English-language text
- Tasks requiring factual knowledge beyond its training and retrieved context
Bias, Risks, and Limitations
The model inherits biases from its base model (Qwen2.5-32B-Instruct) and from the NHS A-to-Z conditions corpus used in the reference application. It may not generalise well to knowledge domains outside its fine-tuning distribution. As with all LLMs, outputs should not be treated as authoritative medical advice.
Recommendations
Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. In particular, outputs from this model in health-related contexts should be reviewed by qualified professionals before being acted upon.
How to Get Started with the Model
1# Use a pipeline as a high-level helper
2from transformers import pipeline
3
4pipe = pipeline("text-generation", model="alan-turing-institute/t0-1.1-k5-32B")
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/t0-1.1-k5-32B")
5model = AutoModelForCausalLM.from_pretrained("alan-turing-institute/t0-1.1-k5-32B")
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]:]))
For serving with vLLM:
1pip install vllm
2vllm serve "alan-turing-institute/t0-1.1-k5-32B"
See the
t0-1 repository and
serve_t0.md for full instructions on setting up the RAG pipeline.
Training Details
Training Data
The reference application uses data scraped from
NHS A-to-Z condition webpages. The data is processed into a JSONL file where each entry contains a
condition_title and
condition_content field. See the
t0-1 repository for more information on the training data and procedure.
Training Procedure
See the
t0-1 repository for full details on the training procedure, including preprocessing, hyperparameters, and evaluation.
Citation
BibTeX:
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}
APA:
Chan, R. S.-Y., Nanni, F., Lazauskas, T., Wood, R., Yong, P., Tarassenko, L., Girolami, M., Geddes, J., & Duncan, A. (2025). Retrieval-augmented reasoning with lean language models. arXiv preprint arXiv:2508.11386.
