Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
V1 proposed storing several language models of one architecture as a shared basis plus small per-model coefficients.It judged rank 2560 (of 3584) "usable" for Qwen2.5-7B models on the evidence of weight and matrix-vector cosines (0.93and 0.92), without running a reconstructed model. This version runs one. We cut every block matrix of Qwen2.5-Coder-7B-Instruct to its own best rank-*k* approximation, the truncated SVD. By theEckart–Young theorem no rank-*k* factorisation, a shared basis included, has less weight error. We wrote the cutmatrices back into the original Q8_0 file and ran the model end to end. At V1's rank 2560 the 7B stays coherent butis measurably worse: - its next token matches the full model's 72 % of the time;- its perplexity on held-out prose doubles (15.5 → 31.4);- HumanEval falls from 145 to 105 of 164 problems (88.4 % → 64.0 %);- it gets a simple question wrong that the full model gets right. The 0.5B model of the same family, cut to the same fraction of its rank, collapses into repetition. All the damage isin the feed-forward (FFN) matrices, which hold 87 % of the block parameters and keep only 89-90 % of their energy atthat rank. The attention matrices can be cut to 2560 almost for free.
No comments yet — start the discussion below.