CollapseTracker — A Time-Lapse Dataset of Progressive Model Collapse

CollapseTracker is a controlled empirical study and dataset documenting progressive model collapse under recursive self‑training: how language model outputs degrade when models are iteratively trained on data that increasingly contains their own generated text. We run a fixed experimental grid over 3 domains (Wikipedia-style factual text, creative fiction, and technical Q&A), 2 models (GPT‑2 and DistilGPT‑2), and 4 synthetic-to-real mixing ratios (1.0, 0.75, 0.5, 0.25) across 11 generations (0–10), producing 24 independent collapse trajectories and 264 per‑generation metric records.

At each generation and condition, we quantify collapse using a consistent metric suite capturing diversity loss (distinct 1–3 grams), self‑similarity increase (Self‑BLEU), distributional drift (KL divergence to the seed distribution), vocabulary shrinkage (vocabulary coverage), rare knowledge loss (rare‑token survival), repetition, and length statistics. The released artifacts include the synthetic generated samples for each track and generation, the aggregated metrics table, and publication-ready plots summarizing collapse trajectories and onset patterns. The original seed datasets used to initialize the experiment are third‑party resources governed by their own licenses and are therefore not redistributed as part of this release; they should be obtained from the original sources for full reproduction