File size: 1,666 Bytes
3f72f87
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: cc-by-4.0
base_model: HPLT/translate-ml-en-v2.0-hplt_opus
library_name: ctranslate2
pipeline_tag: translation
tags: [translation, marian, ctranslate2, int8, ml, en]
---

# translate-windy-hplt-ml-en — ml to en

Bilingual machine translation in **CTranslate2 INT8** for fast CPU inference.
Fine-tuned by **Windstorm Labs** from [`HPLT/translate-ml-en-v2.0-hplt_opus`](https://huggingface.co/HPLT/translate-ml-en-v2.0-hplt_opus).

## Attribution

Derived from [`HPLT/translate-ml-en-v2.0-hplt_opus`](https://huggingface.co/HPLT/translate-ml-en-v2.0-hplt_opus), from the **HPLT** project, licensed under
**CC-BY-4.0**. **Modified by Windstorm Labs.** Attribution is retained as the licence requires.

## What we did

Converted from the upstream Marian release, LoRA fine-tuned on parallel data for this
language pair, merged into the base weights, then quantized to INT8.

| | |
|---|---|
| Rank / alpha | 8 / 16 |
| Learning rate / steps | 2.5e-06 / 50 |
| Target modules | `q_proj`, `v_proj` |
| Seed | 42 |
| Training data | opus100_ml-en (clinic cache), 2,000 sentence pairs |
| Tensors modified | 36 |

## Provenance

```
base model.bin  sha256  209fdd59fbc7640eaec019d172f0d0a7c5b4f1923b02fd9905b33d31309c0185
this model.bin  sha256  116e488fa84df1ac8124b0bc7f556c46b4b9b487365f0c9a56505475086ffa06
```

Verified after INT8 quantization, on the published artifact.

## Evaluation

FLORES-200 devtest, 200 sentences, beam 4, CTranslate2 INT8 on CPU.

| | spBLEU | chrF |
|---|---:|---:|
| this model | **28.97** | 55.08 |

## Notes

- Single language pair: ml to en.
- FLORES-200 is news and encyclopedic prose.
- No human evaluation was performed.