paperkiln-fetch

Drop an arXiv id. Get the paper's architecture back — with receipts.

pip install paperkiln-fetch

paperfetch 1706.03762
# arXiv:1706.03762  Attention Is All You Need
#   d_model         = 512       [d_model = 512]
#   n_layers        = 6         [N = 6 identical layers]
#   n_heads         = 8         [h = 8 parallel attention layers]
#   ...

paperfetch 2302.13971 --emit-hf llama_config.json
# a config.json that transformers.AutoConfig.from_pretrained opens
# directly — with every extracted value's evidence snippet and every
# declared default riding inside it under the "_paperkiln" key

Why this exists

Most transformer papers are deltas over a known skeleton — dims, depth, heads, norm flavor, activation, position encoding. That makes extraction tractable. What nobody else ships is provenance:

• every extracted value carries the evidence snippet it was read from (the sentence, the table row);
• a paper mentions many architectures but uses one — the contribution-vs-mention scorer separates them, and close calls are reported as contested instead of auto-applied;
• fields the sweep cannot resolve are declared unresolved, never silently guessed — and when --emit-hf must fill a default, the default is named in the output;
• explicit rejections ("we choose not to adopt X") veto a candidate outright.

On a 40-paper ground-truth benchmark the scorer runs at grouped AUROC 0.905, abstention-first, with three documented wrong assertions in 92 verdicts (registered and diagnosed in the benchmark, not hidden — two where a paper states its real choice only indirectly, one where a compound flavor name was shadowed by its own substring). The benchmark, its confidence intervals, and its growth protocol ship in the paperkiln repo, where the extracted architectures also train — as real models, on a readable C++ stack, with the paper's actual hyperparameters.

The full loop lives upstream

This package is the extractor, standalone, zero build step. The mother repo adds: constructor-real training of the extracted architecture (--emit-cpp, mtsweep specs), GGUF export, in-browser chat with the model you just trained, and the Architecture Atlas — a pre-registered findings registry for architecture claims.

MIT. Source of truth: papers/fetch.py in the paperkiln repo; this package vendors it verbatim (checked by tools/sync_fetch_pkg.py).