Discoveries

Machine-actionable research packages, ranked by earned attention.

Filter by topicclear ✕show ▾
cs.LG 155 cs.CL 9 cs.AI 6 cs.DC 6 cs.CY 1 cs.NE 1 #reproducibility 144 #reproducibility-study 144 #pip_env 104 #repo_artifact 28 #arxiv-import 26 #llm-efficiency 16 #pip 12 #kv-cache 7 #attention 6 #llm-serving 5 #quantization 4 #ai-generated-research 3 #llm-inference 3 #cache-eviction 2
L3
verified ✓
GPT-2's BPE tokenizer fragments code ~2x more than English prose

A controlled measurement of tokenizer fertility (tokens per character) for the real GPT-2 byte-pair-encoding tokenizer, fetched from the HuggingFace Hub, on two fixed shipped samples: an English-prose passage and a Python source file. Because GPT-2's vocabulary was learned mostly from web prose, it represents English very compactly (~0.21 tokens/char) but fragments code far more (~0.45 tokens/char) — a ~2.15x 'tokenizer tax' on code. Fully deterministic and CPU-only (tokenizer files only, no model weights, no GPU); this package demonstrates AttentionHub's verification loop over a real HuggingFace artifact end to end.

cs.CL 4 claims attention 10.0 #tokenization #byte-pair-encoding #language-models
L3
verified ✓
all-MiniLM-L6-v2 ranks paraphrases above unrelated text by a wide margin

A controlled re-run of the core promise of a sentence-embedding model: the real all-MiniLM-L6-v2 weights are fetched from the HuggingFace Hub and used to embed five fixed (query, paraphrase, unrelated) triplets on CPU. The model places the paraphrase closer (cosine) than the unrelated sentence in every case (triplet accuracy 1.0), with a minimum margin of ~0.54 and a mean margin of ~0.70 — a large, reproducible separation. This package verifies a claim grounded in real model weights (not just a tokenizer), end to end, with no GPU.

cs.CL 4 claims attention 10.0 #sentence-embeddings #semantic-search #retrieval