Create a Reasoning-Focused LLM: A Practical Guide to Streaming, Curating, and Fine-Tuning the SupraLabs Reasoning Corpus

A Practical Guide to Streaming, Curating, and Fine-Tuning the SupraLabs Reasoning Corpus">

In this tutorial, we build an end-to-end workflow for working with the SupraLabs reasoning corpus . We stream a representative subset directly from the Hugging Face Hub, inspect its source distribution, token-length patterns, task composition, and reasoning-to-answer ratios, and then apply a series of quality filters to remove unsuitable training examples. We transform the retained samples into a chat-based supervised fine-tuning format with explicit reasoning tags and use them to adapt SmolLM2-135M-Instruct with LoRA through TRL’s SFTTrainer. By combining scalable data access, exploratory analysis, dataset curation, parameter-efficient fine-tuning, structured inference, and Parquet export, we create a complete Google Colab pipeline for turning a large multi-model reasoning corpus into a compact reasoning-focused language model.

We configure the Colab environment, install the required machine learning libraries, and remove the incompatible torchao package. We detect the available compute device, connect to the SupraLabs reasoning corpus through Hugging Face streaming, and avoid downloading the complete dataset. We shuffle the streamed records, materialize a representative sample, and inspect the structure and contents of an example row.

We convert the sampled dataset into a pandas DataFrame and analyze the distribution of source repositories and token lengths. We calculate reasoning and answer character counts, measure the reasoning-to-response ratio, and visualize the relationships across the dataset. We also apply lightweight heuristic rules to classify each record as a code, mathematics, medical, multiple-choice, or general task.

We construct a quality-filtering pipeline that removes samples with unsuitable token lengths, incomplete responses, excessive repetition, or unbalanced reasoning content. We load the SmolLM2 tokenizer and transform each retained record into a structured conversation containing a system prompt, user message, and reasoning-enhanced assistant response. We then shuffle the formatted data, create training and evaluation subsets, and inspect the final chat template used for supervised fine-tuning.

We load the SmolLM2 causal language model and configure LoRA adapters for parameter-efficient training. We define the optimization, batching, evaluation, precision, and gradient-checkpointing settings through TRL’s SFTConfig. We initialize the SFTTrainer, fine-tune the model on the curated reasoning conversations, and evaluate its final training performance.

We create an inference function that formats new questions with the same system prompt and generates responses from the fine-tuned model. We separate the generated section from the final answer and test the model on logic and arithmetic problems. We finally export the processed training and evaluation datasets as Parquet files for reuse in larger experiments.