falcon-7B Instruct: benchmark results
TII's Falcon-7B-Instruct, a 7B causal decoder-only chat model fine-tuned from Falcon-7B on a mixture of chat and instruct datasets, using FlashAttention and multi-query attention. Provider: TII. Released 2023-05-25. Access: Open.
Aggregate Rating 1152 ± 10, rank #1622 of 1656 rated models, from 126 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| LLM Trustworthy - Fairness | 100 | Trust Score (%) | 93.8 |
| HELM Classic - Entity Data Imputation | 83.59 | Exact Match (%) | 87.9 |
| HELM Classic - Entity Matching | 82.42 | Exact Match (%) | 55.3 |
| MMLU-by-task - TruthfulQA MC1 | 28.89 | Accuracy (%) | 52.3 |
| MMLU-by-task - Machine Learning | 32.14 | Accuracy (%) | 48.9 |
| HELM Classic - LSAT | 20.43 | Exact Match (%) | 48.5 |
| MMLU-by-task - TruthfulQA MC2 | 44.07 | Accuracy (%) | 46.8 |
| HELM Classic - GSM8K | 5.2 | Exact Match (%) | 45.6 |
| MMLU-by-task - Abstract Algebra | 29 | Accuracy (%) | 45.4 |
| MMLU-by-task - Moral Scenarios | 25.14 | Accuracy (%) | 45.4 |
| LLM Trustworthy - Stereotype | 87 | Trust Score (%) | 43.8 |
| Hallucinations Leaderboard | 40.57 | Average Task Score (score) | 42.1 |
Interactive version: theaggregate.ai/model?slug=falcon-7b-instruct · How It Works · Data refreshed daily, snapshot 2026-10-20.