Add README
Browse files
README.md
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
language:
|
| 4 |
+
- ba
|
| 5 |
+
tags:
|
| 6 |
+
- bashkir
|
| 7 |
+
- full-fine-tuning
|
| 8 |
+
- gpt2
|
| 9 |
+
- low-resource
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# Bashkir Full Fine-Tuning Models
|
| 13 |
+
|
| 14 |
+
Full fine-tuned GPT-2 family models on **Bashkir Web Corpus** (71,567 documents).
|
| 15 |
+
|
| 16 |
+
Part of the BashkirNLP project. See the paper for details.
|
| 17 |
+
|
| 18 |
+
## Structure
|
| 19 |
+
|
| 20 |
+
Each top-level subfolder `<model>_seed<N>/` contains a full fine-tuned language model.
|
| 21 |
+
The `classification/` folder contains sequence-classification heads trained on top of the same base models.
|
| 22 |
+
|
| 23 |
+
```
|
| 24 |
+
distilgpt2_seed42/
|
| 25 |
+
distilgpt2_seed123/
|
| 26 |
+
distilgpt2_seed2024/
|
| 27 |
+
gpt2_seed42/
|
| 28 |
+
...
|
| 29 |
+
classification/
|
| 30 |
+
cls_binary_distilgpt2_seed42/
|
| 31 |
+
cls_multiclass_distilgpt2_seed42/
|
| 32 |
+
...
|
| 33 |
+
```
|
| 34 |
+
|
| 35 |
+
## Training
|
| 36 |
+
|
| 37 |
+
- Base models: `distilgpt2` (82M), `gpt2` (124M), `gpt2-medium` (355M)
|
| 38 |
+
- Corpus: Bashkir Web Corpus (57,254 train / 7,157 val / 7,156 test)
|
| 39 |
+
- Epochs: 3
|
| 40 |
+
- Seeds: 42, 123, 2024
|
| 41 |
+
- `max_seq_len`: 128
|
| 42 |
+
|
| 43 |
+
## Evaluation
|
| 44 |
+
|
| 45 |
+
See `step1_report.txt` in the paper repository for metrics (PPL, BPB, catastrophic forgetting, target-language ratio).
|