Skip to main content
Back

Deakin's SPaCe Matches Full-Dataset AI Training With Just 100 Examples

A self-paced curriculum framework matches state-of-the-art reasoning fine-tuning using up to 100x fewer training samples by treating data clusters as bandit arms.

  • SPaCe cuts RL fine-tuning data by up to 100x while matching or beating full-dataset baselines on reasoning benchmarks.
  • Two phases: k-means clustering over embeddings plus difficulty, then Thompson Sampling over clusters as bandit arms.
  • On Qwen3-0.6B it wins all four math benchmarks; on DeepSeek-R1-Distill-1.5B it lifts AIME25 by 10 points.
  • Uses at most 100 training examples total (K under 10 clusters, 10 samples each) with negligible runtime overhead.
  • Stagnating clusters get automatically downweighted so training does not get stuck on unlearnable-yet problems.
  • Built on the Open-R1 codebase and compatible with GRPO out of the box.

Reinforcement fine-tuning has become the default recipe for squeezing reasoning ability out of language models, but it burns through data and GPU hours at a rate that is hard to justify outside big labs. A team from Deakin University proposes a leaner path: a self-paced curriculum framework called SPaCe that picks roughly 100 training examples, adapts what the model sees online, and matches or beats baselines trained on 10,000+ samples.

Standard reinforcement fine-tuning uniformly samples batches from the full dataset, which ignores each example's difficulty, informativeness, and uncertainty, wasting limited reward feedback on trivial or noisy instances and slowing convergence. SPaCe replaces uniform sampling with a two-stage pipeline that first shrinks the training set aggressively, then lets a multi-armed bandit decide which slice of data the model should see at each step.

The two moves behind the method

SPaCe operates in two phases that run back-to-back: a one-time clustering pass over the training set, followed by an online scheduler that runs throughout fine-tuning.

SPaCe architecture showing clustering, data reduction, and bandit selection feeding into RFT
  1. Cluster-based data reduction. Each training example is turned into a joint feature by concatenating a sentence embedding (from Qwen3-Embedding-0.6B, reduced via PCA) with a scalar difficulty score. K-means then partitions the dataset into K clusters, ensuring coverage across the joint latent-attribute space. Within each cluster, farthest-point sampling picks a small, diverse quota of representatives.
  2. Bandit-based data assignment. Each cluster becomes an arm. At every step, Thompson Sampling picks the arm whose current "hardness" score looks most promising, draws a batch from it, updates the policy with GRPO, and refreshes the arm's solve-rate statistics.