File size: 2,953 Bytes
5f0f6b2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: mit
task_categories:
  - text-classification
language:
  - tr
tags:
  - nli
  - turkish
  - error-analysis
  - ensemble
pretty_name: Turkish NLI Wrong Samples (Ensemble Base Models + BiLSTM)
size_categories:
  - 10K<n<100K
configs:
  - config_name: bert-allnli-tr
    data_files: bert-allnli-tr/data.parquet
  - config_name: mdeberta-v3-base-mnli-xnli
    data_files: mdeberta-v3-base-mnli-xnli/data.parquet
  - config_name: gemma-3-27b-it
    data_files: gemma-3-27b-it/data.parquet
  - config_name: qwen2-7b-instruct
    data_files: qwen2-7b-instruct/data.parquet
  - config_name: bilstm-16feat-entropy
    data_files: bilstm-16feat-entropy/data.parquet
---

# Turkish NLI — Wrong Samples (4 Ensemble Bases + BiLSTM)

Wrongly predicted examples for the four ensemble base models and the best
meta-learner (**BiLSTM 16-feature + entropy**), from the Turkish NLI project
(`yilmazzey/sdp2-nli` evaluation splits).

## Prediction source (consistent across all configs)

All **base-model** wrong sets use:

`argmax` over softmax arrays in
`stacking_cache_train_split_16feat_entropy/`

| Model | Softmax origin |
|------|----------------|
| BERT (`emrecan/bert-base-turkish-cased-allnli_tr`) | Encoder classifier softmax |
| mDeBERTa (`MoritzLaurer/mDeBERTa-v3-base-mnli-xnli`) | Encoder classifier softmax |
| Gemma 3 27B IT | **Verbalizer** softmax (last prompt position) |
| Qwen2 7B Instruct | **Verbalizer** softmax (last prompt position) |

**BiLSTM** uses the same cached softmax vectors as 16 features
(12 probabilities + 4 entropies) with the trained `*_bilstm.pt` meta-learner
per dataset config. So base-model wrongs and BiLSTM wrongs are comparable
on the same scoring path (not first-word generation / hard JSON).

## Configs (load one model at a time)

```python
from datasets import load_dataset

ds = load_dataset("yilmazzey/wrong_samples", "bert-allnli-tr")
ds = load_dataset("yilmazzey/wrong_samples", "mdeberta-v3-base-mnli-xnli")
ds = load_dataset("yilmazzey/wrong_samples", "gemma-3-27b-it")
ds = load_dataset("yilmazzey/wrong_samples", "qwen2-7b-instruct")
ds = load_dataset("yilmazzey/wrong_samples", "bilstm-16feat-entropy")
```

## Evaluation splits covered (`split_key`)

| split_key | dataset | split |
|-----------|---------|-------|
| `snli_tr_1_1::test` | snli_tr_1_1 | test |
| `multinli_tr_1_1::validation_matched` | multinli_tr_1_1 | validation_matched |
| `multinli_tr_1_1::validation_mismatched` | multinli_tr_1_1 | validation_mismatched |
| `trglue_mnli::test_matched` | trglue_mnli | test_matched |
| `trglue_mnli::test_mismatched` | trglue_mnli | test_mismatched |

Columns `dataset`, `split`, `split_key`, and `pred_source` are present on every row.

## Notes

- Each row is an example where **that config's model** predicted incorrectly.
- Peer predictions from the other base models are included for recovery analysis.
- This is **not** the Table-4 hard-vote JSON subset (first-word generation path).