GroundTruth

Evaluation and training data for AI systems

[GroundTruth @GroundTruth_IN]

Findings and trajectories: https://t.co/5iJEcamR1c

[GroundTruth @GroundTruth_IN]

Can agents keep up when the right answers change over time?

We tested this with BFSI-Bench: across 9 models, accuracy was 39% on questions with 2026 sources vs 82% with earlier sources.

[Aman Goutam @amansinghgoutam]

Hiring a plant pathology expert at @GroundTruth_IN

Scope: crop image selection, taxonomy, labels, severity criteria, guidelines, reference annotations, team training and quality review.

Paid consulting. DM your expertise. Referrals welcome.

[GroundTruth @GroundTruth_IN]

We're looking for agriculture image annotators!

[Aman Goutam @amansinghgoutam]

@GroundTruth_IN is looking for agriculture image annotators. Agriculture background or relevant annotation experience preferred.

Paid project work. Training provided.

DM your experience and availability.

Referrals welcome.

[GroundTruth @GroundTruth_IN]

Finding the right search result doesn't guarantee a correct answer.

In BFSI Bench, 23% of incorrect web-search agent runs encountered the current regulation but still answered incorrectly.

Trajectories: https://t.co/5iJEcamR1c

[GroundTruth @GroundTruth_IN]

8/ The methodology, results, error analysis, and agent trajectories are available for independent review: https://t.co/5iJEcamR1c
We will periodically update the benchmark to reflect newly published rules and regulatory changes.

[GroundTruth @GroundTruth_IN]

1/ Can frontier AI models reliably interpret India's BFSI regulations, with access to web search? 🧵
India’s BFSI regulations evolve over time. For example, income-tax slabs have changed multiple times in recent years.

[GroundTruth @GroundTruth_IN]

7/ Message limit: Agents exhausted their message budget without reaching the correct answer.
Even with web access, current frontier models struggle to reliably find, interpret, and apply India's BFSI regulations.

[GroundTruth @GroundTruth_IN]

8/ The methodology, results, error analysis, and agent trajectories are available for independent review: https://t.co/5iJEcamR1c
We will periodically update the benchmark to reflect newly published rules and regulatory changes

[GroundTruth @GroundTruth_IN]

7/ Message limit: Agents exhausted their message budget without reaching the correct answer.
Even with web access, current frontier models struggle to reliably find, interpret, and apply India's BFSI regulations.