ANLI
Models label premise and hypothesis pairs as entailment, neutral, or contradiction. People wrote the pairs over three rounds to fool strong models.
Model cards report the rounds R1, R2, and R3 separately or as an average.
- Released
- October 2019
- Kind
- Language reasoning
- Level
- Everyday
- Scored
- Checked automatically
- Answers
- Multiple choice
- Tasks
- All public
- Strongest model tested
-
phi-3-small 7.4B #192 on the capability index - Last in a launch post
- April 2024, dropped before it saturated
- Last model added
- August 2024
- Scores refreshed
- November 12th, 2026
In the authors’ words
We introduce a new large-scale NLI benchmark dataset, collected via an iterative, adversarial human-and-model-in-the-loop procedure. We show that training models on this new dataset leads to state-of-the-art performance on a variety of popular NLI benchmarks, while posing a more difficult challenge with its new test set. Our analysis sheds light on the shortcomings of current state-of-the-art models, and shows that non-expert annotators are successful at finding their weaknesses. The data collection method can be applied in a never-ending learning scenario, becoming a moving target for NLU, rather than a static benchmark that will quickly saturate.
PaLM 2-L leads with 73.1% among models released by April 2024
Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.
PaLM 2-L leads with 73.1%, claims included
Best score per model from every source, including labs’ own numbers and claims Epoch AI collected from reports and papers. Labs pick their own settings, so a claim can beat an independent run of the same model.
The newest model with a score here came out in April 2024, and nobody has reported a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- Claude Fable 5.2
-
Gemini 4 Carbon - GPT-6 Astra
- Claude Opus 5.5
- and 16 more
- 1 PaLM 2-L 73.1%
- 2 phi-3-medium 14B 58.7%
- 3 phi-3-small 7.4B 58.1%
- 4 phi-3-mini 3.8B 52.8%
Source: Labs’ model cards
- 1 PaLM 2-L 73.1%
- 2 phi-3-medium 14B 58.7%
- 3 GPT-3.5 Turbo Nov 2023 58.1%
- 4 phi-3-small 7.4B 58.1%
- 5
Llama 3-8B 57.3% - 6
Mixtral 8x7B 55.2% - 7 phi-3-mini 3.8B 52.8%
- 8 Gemma 7B 48.7%
- 9 Mistral 7B v0.1 47.1%
- 10 Phi-2 42.5%
- 11 Megatron-Turing NLG 530B 39.7%
- 12 InstructGPT 175B 35.4%
Sources: Labs’ model cards and Another board, via Epoch AI (CC BY 4.0)
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | PaLM 2-L | 73.1% | May 2023 | Lab model card | |
| 2 | phi-3-medium 14B | 58.7% | Apr 2024 | Lab model card | |
| 3 | phi-3-small 7.4B | 58.1% | Aug 2024 | Lab model card | |
| 4 | phi-3-mini 3.8B | 52.8% | Aug 2024 | Lab model card |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | PaLM 2-L | 73.1% | May 2023 | Lab model card | |
| 2 | phi-3-medium 14B | 58.7% | Apr 2024 | Lab model card | |
| 3 | GPT-3.5 Turbo Nov 2023 | 58.1% | Nov 2023 | Claim in Phi-3 Technical Report: A Highly Capable Language Model L…, via Epoch AI | |
| 4 | phi-3-small 7.4B | 58.1% | Aug 2024 | Lab model card | |
| 5 | Llama 3-8B | 57.3% | Apr 2024 | Claim in Phi-3 Technical Report: A Highly Capable Language Model L…, via Epoch AI | |
| 6 | Mixtral 8x7B | 55.2% | Dec 2023 | Claim in Phi-3 Technical Report: A Highly Capable Language Model L…, via Epoch AI | |
| 7 | phi-3-mini 3.8B | 52.8% | Aug 2024 | Lab model card | |
| 8 | Gemma 7B | 48.7% | Feb 2024 | Claim in Phi-3 Technical Report: A Highly Capable Language Model L…, via Epoch AI | |
| 9 | Mistral 7B v0.1 | 47.1% | Sep 2023 | Claim in Phi-3 Technical Report: A Highly Capable Language Model L…, via Epoch AI | |
| 10 | Phi-2 | 42.5% | Dec 2023 | Claim in Phi-3 Technical Report: A Highly Capable Language Model L…, via Epoch AI | |
| 11 | Megatron-Turing NLG 530B | 39.7% | Oct 2021 | Claim in Using DeepSpeed and Megatron to Train Megatron-Turing NLG…, via Epoch AI | |
| 12 | InstructGPT 175B | 35.4% | Jan 2022 | Claim in Using DeepSpeed and Megatron to Train Megatron-Turing NLG…, via Epoch AI |
Epoch AI collected claims for 11 models from labs’ reports and papers. Switch the scores above to rank them together.
Number one over time 2 models have held number one with claims included since 2023
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Labs’ model cards
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Sources: Labs’ model cards and claims collected by Epoch AI (CC BY 4.0)
No launch post has listed it since April 2024
Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.
| Date | Launch post | Score | As the post writes it |
|---|---|---|---|
| Apr 22, 2024 | phi-3-medium 14B | 58.7 | ANLI (7-Shot) |
| May 10, 2023 | PaLM 2-L | 73.1 | Best of ANLI-R1, 1-shot / ANLI-R2, 1-shot / ANLI-R3, 1-shot |
3 system cards report it, from 1 lab
Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| Aug 30, 2024 | phi-3-small 7.4B technical report | 58.1 | ANLI (7-Shot) (Section 3 table) | Capabilities |
| Aug 30, 2024 | phi-3-mini 3.8B technical report | 52.8 | ANLI (7-Shot) (Section 3 table) | Capabilities |
| Aug 30, 2024 | phi-3-medium 14B technical report | 55.8 | ANLI (7-Shot) (Section 3 table) | Capabilities |
More in Knowledge
- MMLU-Pro In use Harder 10-option version of MMLU with reasoning-focused questions across 14 subjects.
Intern-S2-397B 89.8% - SuperGPQA In use 26,529 graduate-level multiple-choice questions across 285 disciplines. Accuracy.
Qwen 3.6 Max (Preview) 73.9% - WideSearch In use 200 bilingual broad info-seeking tasks: agent collects many atomic facts into a table. Scored by success rate and row/item F1.
Tencent Hy4 preview 83.9% - BBEH micro average In use Models solve harder replacements for each of the 23 BIG-Bench Hard tasks, 4,520 examples in total. The score is the share of all examples answered correctly. Gemma 4 31B IT 74.4%
- KMMLU-Pro In use Models answer multiple-choice questions taken from Korean National Professional Licensure exams. It tests professional knowledge used in Korean industry.
Solar Pro 4 79.2% - HLE-Verified In use Verified and revised version of Humanity's Last Exam (1,811 items) with uncertain items removed.
Seed2.0 Pro 73.6%