File size: 4,881 Bytes
3a225e6
9b44268
3a225e6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b18dbbb
 
 
 
 
198b08b
3a225e6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b18dbbb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1191df3
3a225e6
1191df3
 
 
 
 
3a225e6
 
 
32a09f9
1191df3
 
 
 
 
 
 
 
 
 
 
 
 
 
32a09f9
3a225e6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
---
license: cc-by-4.0
language:
- es
task_categories:
- text-classification
task_ids:
- natural-language-inference
pretty_name: ESNLIR  test split
size_categories:
- 10K<n<100K
tags:
- nli
- spanish
- causal
- esnlir
configs:
- config_name: default
  data_files:
  - split: test
    path: test.json
---

# ESNLIR — test split

The full test split of **ESNLIR**, a Spanish multi-genre NLI dataset with a *reasoning* (causal)
label: **80,213 sentence pairs**, label-balanced by construction.

This is the exact test split used to evaluate GPT-4o, GPT-4o-Mini, Qwen2.5-7B-Instruct and
Llama-3.1-8B-Instruct in **An Analysis of the Performance of Large Language Models in Spanish NLI
Datasets with Causal Relationships** (IBERAMIA 2026, to appear); code in
[`Pacolas/NLI-via-LLM`](https://github.com/Pacolas/NLI-via-LLM). The data itself comes from ESNLIR,
a separate contribution with its own code and paper. Part of the
[**ESNLIR-LLM**](https://huggingface.co/collections/Flaglab/esnlir-llm-6a74a6b9ce149918f48ebba5) collection.

## Format

Despite the `.json` extension, the file is **JSON Lines** — one object per line, no enclosing
brackets.

```json
{"sentence_1": "...", "sentence_2": "...", "connector": "por eso",
 "connector_type": "reasoning", "extraction_strategy": "linking_phrase", "distance": 1.0,
 "sentence_1_paragraph": 46, "sentence_1_position": 25,
 "sentence_2_paragraph": 46, "sentence_2_position": 26,
 "id": "esnews__spanish_pd_news__365259", "dataset": "esnews__spanish_pd_news",
 "genre": "news", "domain": "spanish_public_domain_news"}
```

The gold label is **`connector_type`**, derived from the discourse connector linking the two
sentences — it is *not* human-validated. For the human-validated subset see
[`Flaglab/esnlir-human-validated`](https://huggingface.co/datasets/Flaglab/esnlir-human-validated).

## Label distribution

| label | n |
|---|---|
| `entailment` | 20,054 |
| `contrasting` | 20,054 |
| `reasoning` | 20,053 |
| `neutral` | 20,052 |

## Genre distribution

| genre | n |
|---|---|
| `theses` | 28,348 |
| `books` | 17,540 |
| `comments` | 11,137 |
| `news` | 10,248 |
| `articles` | 6,568 |
| `legal` | 5,628 |
| `talks` | 548 |
| `clinical` | 196 |

## Usage

```python
from huggingface_hub import hf_hub_download
import json

p = hf_hub_download("Flaglab/esnlir-test", "test.json", repo_type="dataset")
rows = [json.loads(l) for l in open(p) if l.strip()]
```

## Reported results on this split

From the IBERAMIA 2026 paper (accuracy / macro F1), all LLMs using `prompt_0004` with Few-Shot and
shared context, no fine-tuning:

| model | setting | accuracy | macro F1 |
|---|---|---|---|
| XLM-RoBERTa | fine-tuned | **0.676** | **0.676** |
| BERTIN | fine-tuned | 0.663 | 0.664 |
| Qwen2.5-7B-Instruct | few-shot | 0.411 | 0.370 |
| GPT-4o-Mini | few-shot | 0.390 | 0.386 |
| XGBoost | fine-tuned | 0.350 | 0.348 |
| Llama-3.1-8B-Instruct | few-shot | 0.326 | 0.296 |
| Majority class | — | 0.250 | 0.250 |

A fine-tuned encoder beats every prompted 7–8B model by more than 25 points. Note also that both
open models nearly abandon Entailment (≈0.08 recall), effectively collapsing the four-way task into
a three-way one.

## Caveat when comparing runs

vLLM applies no content filtering, so open models served locally return a label for all 80,213
pairs. The GPT-4o / GPT-4o-Mini runs went through Azure, whose NSFW filter rejected 44 pairs (40 of
them Neutral). The two bases are therefore not exactly identical, and the missing pairs concentrate
in the sensitive-content source datasets — worth keeping in mind when comparing proprietary and
open-weight numbers pair for pair.

## Citation

These splits were packaged for the following paper:

```bibtex
@InProceedings{perez2026llmspanishnlicausal,
  author    = {P{\'e}rez, Nicol{\'a}s and Portela, Johan R. and Manrique, Ruben},
  title     = {An Analysis of the Performance of Large Language Models in Spanish
               NLI Datasets with Causal Relationships},
  booktitle = {Advances in Artificial Intelligence -- IBERAMIA 2026},
  year      = {2026},
  publisher = {Springer Nature Switzerland},
  address   = {Cham},
  note      = {To appear},
}
```

The data itself is ESNLIR, released here under CC BY 4.0 — attribution to the source corpus is a condition of that licence:

```bibtex
@InProceedings{portela2025esnlirspanishmultigenredataset,
  author    = {Portela, Johan R. and P{\'e}rez-Ter{\'a}n, Nicol{\'a}s and Manrique, Rub{\'e}n},
  editor    = {Florez, Hector and Peluffo-Ordo{\~{n}}ez, Diego},
  title     = {{ESNLIR}: Expanding Spanish {NLI} Benchmarks with Multi-genre and Causal Annotation},
  booktitle = {Applied Informatics},
  year      = {2026},
  publisher = {Springer Nature Switzerland},
  address   = {Cham},
  pages     = {345--361},
  isbn      = {978-3-032-07175-0},
  doi       = {10.1007/978-3-032-07175-0_23},
}
```