Dev claim · claim-897 · #141 of 154
“The report […] found that the United States was one of the most pollution-free nations in the world.””
Claim tags (sorted stems; highlighted when a passage shares them)
- found
- free
- nation
- one
- pollut
- report
- state
- unit
- world
The model’s verdict
The Transformer retrained from the notebook (the 2024 weights were never saved), shown three ways. The first is how the notebook evaluated it.
Notebook protocol
Dev claims predicted in batches of 16, in file order, from the 2024 retrieved evidence, as the notebook does.
- Supports
- 59.7%
- Refutes
- 34.3%
- Not enough info
- 0.4%
- Disputed
- 5.6%
One claim at a time
The same model run on a single claim, which is how the Try-it page runs it.
- Supports
- 53.2%
- Refutes
- 37.7%
- Not enough info
- 5.5%
- Disputed
- 3.7%
With gold evidence
Fed the human-annotated evidence instead of retrieved passages. This is the 'val accuracy' the training loop reports.
- Supports
- 62.2%
- Refutes
- 30.4%
- Not enough info
- 0.5%
- Disputed
- 6.9%
Retrieved vs gold evidence
Gold passages were picked by the dataset’s annotators. Retrieved passages are what the TF-IDF rule returned from all 1.19M passages. Switch between the submitted 2024 output and the re-runs.
2024 submission
The passages the team actually retrieved and submitted in 2024 (their saved output file). The file records only passage ids, so scores are recomputed with the submission rule.
- Path
- not recorded
- Found
- 0 of 1 gold
- P
- 0.00
- R
- 0.00
- F
- 0.00
#1evidence-969926 According to the United Nations' World Population Prospects report:
- accord
- report (shared with the claim)
- unit (shared with the claim)
- nation (shared with the claim)
- world (shared with the claim)
- popul
- prospect
- cos
- 0.582
- overlap
- 0.571
- score
- 1.153
- shared
- 4
Gold evidence (1)
evidence-783786 Since 2007, the total greenhouse gas emissions by the United States are the second highest by country, exceeded only by China.
- exceed
- china
- highest
- emiss
- ga
- greenhou
Optional · uses your own API key
Second opinion from an LLM
The model reads this claim and the evidence passages shown on this page, picks one of the four verdicts and cites the passages it used. It is told to judge from the passages only. You review the answer; the call and your decision go to the AI audit log in this browser.
Without a key nothing is sent anywhere. The rest of the site does not need one.