TPBench measures three retrieval targets after dialogue compression: This repository includes data, construction inputs, compression and reader code, scorers, saved answers, and paper results. RESULTSMAP.md connects the paper comparisons to their evidence files. Use Python 3.10 or newer. No model access is needed for these checks. The checks validate file hashes and metadata, regenerate all 42 main probe cells, rescore saved answers, and test FMTS context construction. Outputs from tests are written to a temporary directory, not over the released results. Each row fixes the example identifier, method, budget, reference answer, and reader prompts. Identifiers remain stable across…
Independent publisher
Papersubmission
4papersubmission
Models in Library0
Datasets in Library1
Models on Hugging Face—
Followers—
Compare the published evidence
This profile connects the 1 model and dataset record shown below to their upstream cards and sources. The library counts describe SAVRN's coverage, rather than this publisher's entire output.
Published tasks: Question answering.
Reported licenses: other. Open a record to review its specific license and conditions.
Compare model specifications, source dates, and hardware estimates on the individual records. Publisher identity verification is separate from a model's quality, deployment readiness, or suitability for your work.