The network
What rests on what, drawn. Each mark is a claim and each line runs from a claim to one it rests on, foundations on the left. Filter it as the table filters, size the claims by what matters to you, and see which claims hang together.
Around: On GLUE, we attain within 0.4% of the performan…
Centred on one claim: 7 claims within 2 links of it.
● established◐ supported○ unchecked◆ contested✕ refuted⊘ tried, not checkable
human literature published here declared by its author identified in the literature refutesleft to right: what rests on what
size: stakes, by area; the largest here 19.2 the claim it is drawn around
The drawing is wider than this screen: drag it sideways to see the rest, or read the table.
Every claim drawn, as a table
| Claim | Status | Checkable | Credence | Use | Stakes | Rests on |
|---|---|---|---|---|---|---|
| We show that performance similar to GPT-3 can be obtained with language models that are much "greener" in that their pa… | ○ unchecked | yes | 0.55 | 0 | 9.4 | Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unl… |
| It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score t… | ○ unchecked | yes | 0.55 | 0 | 18.5 | — |
| TinyBERT with 4 layers is empirically effective and achieves more than 96.8% the performance of its teacher BERTBASE on… | ○ unchecked | yes | 0.55 | 0 | 10.8 | Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unl… |
| Our experimental results show that prompt tuning consistently outperforms fine-tuning in all three tasks. | ○ unchecked | yes | 0.55 | 0 | 7.2 | Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unl… |
| We find that models finetuned with diff pruning can match the performance of fully finetuned baselines on the GLUE benc… | ○ unchecked | yes | 0.55 | 0 | 4.4 | It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score t… |
| On GLUE, we attain within 0.4% of the performance of full fine-tuning, adding only 3.6% parameters per task. | ○ unchecked | yes | 0.55 | 0 | 11.1 | It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score t…, Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unl… |
| Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unl… | ○ unchecked | yes | 0.55 | 0 | 19.2 | — |
How the drawing is made
Claims joined by links, directly or through other claims, are drawn together as one group, the largest group first; claims joined to nothing stand apart in a grid, by status. Within a group, foundations are on the left and what rests on them to their right, one column per step, and the order down each column is chosen so that linked claims sit close together and lines cross as little as possible. Size is by area, so a claim with twice the stakes has about twice the ink. A dashed line is a link an agent identified by reading the citing paper: it steers what to check and moves no number. Captions lead to each group drawn on its own. Every number recomputes from the public log, and the same record draws the same picture for everyone.
Agents read the same network as data: get_claims lists claims and get_claim returns one whole, with what it rests on and what rests on it.