File size: 3,990 Bytes
5eb10b3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: apache-2.0
task_categories:
  - text-classification
language:
  - multilingual
tags:
  - language-identification
  - unigram
  - tokenizer
  - tinyaya
pretty_name: TinyAya LID Experiment Logs
---

# TinyAya LID — Models, Eval Data & Training Artifacts

Artifacts for the **Contrastive UniLID** project: language identification using LLM tokenizer vocabularies (TinyAya 261k BPE→Unigram), trained on GlotLID-C, evaluated on CommonLID.

Source code: [github.com/divyanshsinghvi/tinyAyaLid](https://github.com/divyanshsinghvi/tinyAyaLid)

> **Note**: GlotLID-C training corpus is **not included** here — it can be re-downloaded from [`cis-lmu/glotlid-corpus`](https://huggingface.co/datasets/cis-lmu/glotlid-corpus). This repo only ships the eval data, models, training weights, and LLM cache.

---

## Structure

```
.
├── models/                          # Trained .unilid model files + eval JSONs
│   ├── tinyaya_v3_200k/             # Best TinyAya model — 200k samples/lang
│   ├── tinyaya_v3_100k/             # TinyAya, 100k samples/lang
│   ├── tinyaya_soft_full/           # TinyAya, full GlotLID-C corpus
│   ├── mistral_v3_200k/             # Mistral-Nemo 131k tokenizer comparison
│   ├── scratch_v3_200k/             # Scratch 100k vocab comparison
│   ├── commonlid_20pct/             # Trained on 20% CommonLID split (TinyAya)
│   ├── commonlid_50pct/             # Trained on 50% CommonLID split (TinyAya)
│   ├── commonlid_20pct_mistral/     # 20% CommonLID split (Mistral)
│   ├── commonlid_50pct_mistral/     # 50% CommonLID split (Mistral)
│   ├── commonlid_20pct_scratch/     # 20% CommonLID split (Scratch)
│   └── commonlid_50pct_scratch/     # 50% CommonLID split (Scratch)

├── data/
│   ├── commonlid/                   # CommonLID evaluation corpus (fastText format)
│   │   ├── commonlid_full.txt       # Full test set (373k samples, 109 tags)
│   │   ├── commonlid_train.txt      # Train split
│   │   ├── commonlid_test.txt       # Test split
│   │   ├── commonlid_50pct_test.txt # 50% split
│   │   ├── commonlid_80pct_test.txt # 80% split
│   │   ├── commonlid_50perlang.txt  # 50 samples/lang subsample
│   │   ├── commonlid_150perlang.txt # 150 samples/lang subsample
│   │   ├── commonlid_200perlang.txt # 200 samples/lang subsample
│   │   ├── commonlid_20pct_by_lang/ # Per-language files (20pct split)
│   │   └── commonlid_50pct_by_lang/ # Per-language files (50pct split)
│   │
│   └── misc/                        # Small training experiment files
│       ├── train_quick.txt
│       ├── train_quick_test.txt
│       ├── train_1k.txt
│       ├── train_1k_test.txt
│       └── train_test.txt

├── training_weights/                # Per-language unigram log-prob dists from soft EM (compressed)
│   └── *.tar.gz                     # One tarball per experiment config

└── cache/                           # Cached LLM API responses (two-stage eval)
    └── cache.tar.gz
```

## Data Formats

- **fastText format** (`__label__<lang_Script> <text>`): all CommonLID files
- **Plain text** (one sentence per line): misc training files

## Languages

- **CommonLID eval**: 109 language tags (373,230 samples in `commonlid_full.txt`)
- **Alias mapping** (CommonLID→model individual code):
  `ara→arb, aze→azj, bik→bcl, est→ekk, lav→lvs, mlg→plt, msa→zsm, orm→gaz, swa→swh, tgl→fil, uzb→uzn, zho→cmn`

## Reproducing Training

To retrain a model, download GlotLID-C separately:
```python
from datasets import load_dataset
ds = load_dataset("cis-lmu/glotlid-corpus")
```
Then run `train.py` from the source repo using the desired tokenizer.

## Contributors

Divyansh Singhvi, Megha Agarwal. Mentored by Julia Kreutzer.