---
title: "Dolorem ipsum quia dolor sit amet · YecoAI Research"
description: "Dolorem ipsum quia dolor sit amet, from YecoAI Research: methods, results and open questions on building dependable, safe and efficient language models."
url: "https://yecoai-9487cf1e.deplo.site/en/research/dolorem-ipsum-quia-dolor-sit-amet"
locale: "en"
published: "2025-12-25"
updated: "2026-09-24"
category: "Safety"
---

# Dolorem ipsum quia dolor sit amet

Maecenas sed diam eget risus varius blandit sit amet non magna.

*25 Dec 2025 · Safety*

## Abstract

Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.

Lorem ipsum dolor sit amet consectetur adipiscing elit. Quisque faucibus ex sapien vitae pellentesque sem placerat. In id cursus mi pretium tellus duis convallis.

- **Benchmark accuracy:** 92.4% (+4.1%). Above the previous state of the art
- **Tokens trained:** 3.2T (+18%). Curated pre-training corpus
- **Inference latency:** 118 ms (-23.5%, lower is better). p50 on a single accelerator
- **Hallucination rate:** 2.1% (-41%, lower is better). Measured on closed-book QA

## Training

Lorem ipsum dolor sit amet consectetur. Posuere est id faucibus magnis pharetra nunc tempor netus. Pharetra quis diam auctor aliquam blandit in sit imperdiet. Id quam vitae vel non condimentum lectus.

Lorem ipsum dolor sit amet consectetur adipiscing elit. Quisque faucibus ex sapien vitae pellentesque sem placerat. In id cursus mi pretium tellus duis convallis.

### Accuracy during training

Train and held-out eval accuracy (%) over 100k steps

|  | Train | Eval |
|---|---|---|
| 0k | 12 | 10 |
| 10k | 34 | 29 |
| 20k | 51 | 45 |
| 30k | 63 | 57 |
| 40k | 71 | 66 |
| 50k | 77 | 72 |
| 60k | 82 | 77 |
| 70k | 86 | 81 |
| 80k | 89 | 84 |
| 90k | 91 | 86 |
| 100k | 92 | 88 |

Eval accuracy tracks train within 4 points after 50k steps.

## Scaling across releases

Lorem ipsum dolor sit amet consectetur. Posuere est id faucibus magnis pharetra nunc tempor netus. Pharetra quis diam auctor aliquam blandit in sit imperdiet. Id quam vitae vel non condimentum lectus.

### Score across releases

Aggregate benchmark score per model version

|  | YecoAI | Baseline |
|---|---|---|
| v1 | 61 | 58 |
| v2 | 68 | 62 |
| v3 | 74 | 66 |
| v4 | 81 | 70 |
| v5 | 87 | 73 |
| v6 | 92 | 75 |

The gap to the baseline widens with every release.

## Benchmarks

Lorem ipsum dolor sit amet consectetur. Posuere est id faucibus magnis pharetra nunc tempor netus. Pharetra quis diam auctor aliquam blandit in sit imperdiet. Id quam vitae vel non condimentum lectus.

### Benchmark comparison

Accuracy (%) on five public benchmarks

|  | YecoAI | Model A | Model B |
|---|---|---|---|
| MMLU | 89 | 84 | 79 |
| GSM8K | 94 | 88 | 81 |
| HumanEval | 86 | 80 | 72 |
| ARC | 91 | 87 | 83 |
| HellaSwag | 95 | 93 | 90 |

Best on every benchmark, largest margin on HumanEval.

## Capabilities

Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.

### Capability profile

Normalised scores per capability area

|  | YecoAI | Baseline |
|---|---|---|
| Reasoning | 92 | 78 |
| Coding | 86 | 74 |
| Math | 94 | 80 |
| Knowledge | 89 | 84 |
| Safety | 96 | 82 |
| Multilingual | 83 | 71 |

### Overall score

Weighted average across all evaluations

**87** / 100 out of 100

## Training data

Lorem ipsum dolor sit amet consectetur. Posuere est id faucibus magnis pharetra nunc tempor netus. Pharetra quis diam auctor aliquam blandit in sit imperdiet. Id quam vitae vel non condimentum lectus.

Lorem ipsum dolor sit amet consectetur adipiscing elit. Quisque faucibus ex sapien vitae pellentesque sem placerat. In id cursus mi pretium tellus duis convallis.

### Training data mix

Share of pre-training tokens by source

|  | tokens |
|---|---|
| Web | 1480 |
| Code | 720 |
| Papers | 480 |
| Books | 320 |
| Synthetic | 200 |

Total: 3.2T

## Efficiency

Lorem ipsum dolor sit amet consectetur. Posuere est id faucibus magnis pharetra nunc tempor netus. Pharetra quis diam auctor aliquam blandit in sit imperdiet. Id quam vitae vel non condimentum lectus.

### Inference latency

p50 latency per request, lower is better

|  | Latency (ms) |
|---|---|
| YecoAI | 118 |
| Model A | 154 |
| Model B | 189 |
| Model C | 231 |

## Full results

Lorem ipsum dolor sit amet consectetur adipiscing elit. Quisque faucibus ex sapien vitae pellentesque sem placerat. In id cursus mi pretium tellus duis convallis.

### Full results

Accuracy (%) per model and benchmark

| Model | MMLU | GSM8K | HumanEval | ARC | Avg |
|---|---|---|---|---|---|
| YecoAI | 89 | 94 | 86 | 91 | 90 |
| Model A | 84 | 88 | 80 | 87 | 84.8 |
| Model B | 79 | 81 | 72 | 83 | 78.8 |
| Model C | 76 | 77 | 68 | 80 | 75.3 |

## Conclusion

Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.

Lorem ipsum dolor sit amet consectetur. Posuere est id faucibus magnis pharetra nunc tempor netus. Pharetra quis diam auctor aliquam blandit in sit imperdiet. Id quam vitae vel non condimentum lectus.

```json
{
  "@context": "https://schema.org",
  "@id": "https://yecoai-9487cf1e.deplo.site/en/research/dolorem-ipsum-quia-dolor-sit-amet#page",
  "url": "https://yecoai-9487cf1e.deplo.site/en/research/dolorem-ipsum-quia-dolor-sit-amet",
  "name": "Dolorem ipsum quia dolor sit amet · YecoAI Research",
  "description": "Dolorem ipsum quia dolor sit amet, from YecoAI Research: methods, results and open questions on building dependable, safe and efficient language models.",
  "inLanguage": "en",
  "isPartOf": {
    "@id": "https://yecoai-9487cf1e.deplo.site/#website"
  },
  "@type": "ScholarlyArticle",
  "headline": "Dolorem ipsum quia dolor sit amet",
  "mainEntityOfPage": "https://yecoai-9487cf1e.deplo.site/en/research/dolorem-ipsum-quia-dolor-sit-amet",
  "datePublished": "2025-12-25T00:00:00.000Z",
  "dateModified": "2026-09-24T15:24:11.967Z",
  "articleSection": "Safety",
  "keywords": "YecoAI Research, Safety, Dolorem ipsum quia dolor sit amet, AI research, LLM safety, YecoAI",
  "author": {
    "@id": "https://yecoai-9487cf1e.deplo.site/#organization"
  },
  "publisher": {
    "@id": "https://yecoai-9487cf1e.deplo.site/#organization"
  }
}
```

```json
{
  "@context": "https://schema.org",
  "@type": "BreadcrumbList",
  "itemListElement": [
    {
      "@type": "ListItem",
      "position": 1,
      "name": "YecoAI home",
      "item": "https://yecoai-9487cf1e.deplo.site/en"
    },
    {
      "@type": "ListItem",
      "position": 2,
      "name": "Research",
      "item": "https://yecoai-9487cf1e.deplo.site/en/research"
    },
    {
      "@type": "ListItem",
      "position": 3,
      "name": "Dolorem ipsum quia dolor sit amet",
      "item": "https://yecoai-9487cf1e.deplo.site/en/research/dolorem-ipsum-quia-dolor-sit-amet"
    }
  ]
}
```