ratishsp/g-vendi-base-proxy

Use a base-model proxy for G-Vendi, and score it within format

Companion data for the FinePhrase Synthetic Data Playbook (Niklaus et al., 2026), on the 83-cell grid from its Figure 22. G-Vendi (Jung et al., 2025) is a gradient-space diversity metric that the playbook uses as one of several predictors of downstream performance.

The heatmap correlates each predictor with each benchmark across the grid. The top three blocks are the playbook's other predictors (DCLM, Edu, and the embedding-based diversity metrics: Vendi Score, cosine similarity, near-duplicate rate). The bottom block is G-Vendi, the gradient-based diversity metric, built up one row at a time:

• Published / Reproduced. The first row is the playbook's published G-Vendi; the second is our re-run of the same pipeline from scratch. They agree (cell-level rank-correlation 0.95), so the reproduction is faithful.
• Base model. The third row swaps the proxy model from the instruction-tuned Qwen3-0.6B to its pretrained base, holding everything else fixed. G-Vendi backpropagates each document through a small proxy LLM; it was introduced for curating reasoning datasets, where an instruction-tuned proxy is the natural choice, and the playbook uses one too. We are evaluating pretraining data, so we match the proxy to the data and use the base model.
• Base model, within-prompt. The last row scores the base model within each format (z-scoring inside each (bucket, prompt) group before correlating). This strips out differences between formats, so it tests whether the metric picks the better rephraser for a given format.

In that final row the base-model proxy reaches +0.57 on the macro average (p < 0.001).

Comparing that row against DCLM-difference, the strongest reference predictor: DCLM-difference leads on the macro and micro averages, and math and table; the base-model within-prompt row leads on general knowledge, and on reasoning and NLU, where DCLM has little signal. On reading comprehension the two rows are tied, but the base-model within-prompt row leads on SQuAD v2. Note the two rows are scored differently: DCLM-difference is raw-pooled, while the base-model row is within-prompt.

Files and reproducing

python gvendi_correlations.py   # raw and within-prompt correlations
python make_heatmap.py          # writes correlation_heatmap.png

data/ holds per-cell G-Vendi for the published pipeline (pub.json), our reproduction (repro/), and the base-model swap (prx06bb/), plus the playbook's predictor and downstream scores (reference_predictors.json, downstream_scores.json), both taken from its rephrasing_metadata.json so the raw rows reproduce Figure 22 exactly.

Citation

@misc{puduppully2026gvendiproxy,
  author       = {Puduppully, Ratish},
  title        = {Use a Base-Model Proxy for G-Vendi, and Score It Within Format},
  year         = {2026},
  howpublished = {HuggingFace dataset},
  url          = {https://huggingface.co/datasets/ratishsp/g-vendi-base-proxy}
}

Please also cite the work this builds on: the FinePhrase Synthetic Data Playbook (Niklaus et al., 2026), Prismatic Synthesis / G-Vendi (Jung et al., 2025, arXiv:2505.20161), and the Vendi Score (Friedman & Dieng, 2023).