amr-mohamed
commited on
Commit
•
4626cbb
1
Parent(s):
1ed7b0f
Update README.md
Browse files
README.md
CHANGED
@@ -311,446 +311,6 @@ Our training dataset [Darija-SFT-Mixture](https://huggingface.co/datasets/MBZUAI
|
|
311 |
Atlas-Chat models are based on Gemma 2 models. The Atlas-Chat models were trained using 8 Nvidia's A100 80 GB GPUs in parallel using FSDP on AWS Sagemaker. The model is trained using HuggingFace transformers and parameter-efficient fine-tuning with LoRA rank of 256.
|
312 |
|
313 |
|
314 |
-
<!--
|
315 |
-
## Evaluation
|
316 |
-
The Atlas-Chat models were evaluated on a comprehensive suite of tasks using various datasets and benchmarks to assess their performance across multiple dimensions. These included tasks such as:
|
317 |
-
|
318 |
-
* **DarijaMMLU:** A Darija version of ArabicMMLU and MMLU benchmarks translated from MSA and English respectively.
|
319 |
-
* **DarijaHellaSwag:** A Darija version of HellaSwag.
|
320 |
-
* **Belebele Ary_Arab:** Belebele is a multiple-choice machine reading comprehension dataset published by Facebook spanning 122 language variants. The Evaluation is done on the Ary_Arab part of Belebele that refers to Darija.
|
321 |
-
* **Sentiment Analysis.**
|
322 |
-
* **Translation:** Including six directions and four languages: Darija, MSA, English and French.
|
323 |
-
* **Transliteration:** Transforming a sentence from Darija (written in Arabic characters) to Arabizi (Written in Latin characters) and vice-versa.
|
324 |
-
* **Summarization.**
|
325 |
-
|
326 |
-
The models were compared against a collection of existing open-source Arabic models to gauge their effectiveness, with a particular focus on performance in Darija. All scores are based on zero-shot performance. The prompts are written mainly in Darija. The metric used for DarijaMMLU, DarijaHellaSwag, Belebele Ary and Sentiment Analysis is the normalized accuracy. We used [Language Model Evaluation Harness](https://github.com/MBZUAI-Paris/lm-evaluation-harness-atlas-chat) to conduct these evaluations.
|
327 |
-
|
328 |
-
|
329 |
-
**LLMs Benchmarks:**
|
330 |
-
<table>
|
331 |
-
<tr>
|
332 |
-
<td>Model</td>
|
333 |
-
<td><a href="https://huggingface.co/datasets/MBZUAI-Paris/DarijaMMLU" target="_blank">DarijaMMLU</a></td>
|
334 |
-
<td><a href="MBZUAI-Paris/DarijaHellaSwag" target="_blank">DarijaHellaSwag</a></td>
|
335 |
-
<td ><a href="https://huggingface.co/datasets/facebook/belebele/viewer/ary_Arab" target="_blank">Belebele Ary</a></td>
|
336 |
-
</tr>
|
337 |
-
<tr>
|
338 |
-
<td><a href="https://huggingface.co/inceptionai/jais-family-1p3b-chat" target="_blank">jais-family-1p3b-chat</a></td>
|
339 |
-
<td>35.39</td>
|
340 |
-
<td>32.51</td>
|
341 |
-
<td>38.33</td>
|
342 |
-
</tr>
|
343 |
-
<tr>
|
344 |
-
<td><a href="https://huggingface.co/inceptionai/jais-family-2p7b-chat" target="_blank">jais-family-2p7b-chat</a></td>
|
345 |
-
<td>37.44</td>
|
346 |
-
<td>34.49</td>
|
347 |
-
<td>44.11</td>
|
348 |
-
</tr>
|
349 |
-
<tr>
|
350 |
-
<td><a href="https://huggingface.co/google/gemma-2-2b-it" target="_blank">gemma-2-2b-it</a></td>
|
351 |
-
<td>28.58</td>
|
352 |
-
<td>32.42</td>
|
353 |
-
<td>25.22</td>
|
354 |
-
</tr>
|
355 |
-
<tr>
|
356 |
-
<td><a href="meta-llama/Llama-3.2-1B-Instruct" target="_blank">Llama-3.2-1B-Instruct</a></td>
|
357 |
-
<td>27.66</td>
|
358 |
-
<td>26.88</td>
|
359 |
-
<td>28.89</td>
|
360 |
-
</tr>
|
361 |
-
<tr>
|
362 |
-
<td><a href="meta-llama/Llama-3.2-3B-Instruct" target="_blank">Llama-3.2-3B-Instruct</a></td>
|
363 |
-
<td>32.60</td>
|
364 |
-
<td>28.33</td>
|
365 |
-
<td>38.00</td>
|
366 |
-
</tr>
|
367 |
-
<tr>
|
368 |
-
<td><strong><a href="https://huggingface.co/MBZUAI-Paris/Atlas-Chat-2B" target="_blank">Atlas-Chat-2B</a></strong></td>
|
369 |
-
<td><b>44.97</td>
|
370 |
-
<td><b>41.48</td>
|
371 |
-
<td><b>53.89</td>
|
372 |
-
</tr>
|
373 |
-
<tr style="border-top: 4px solid;"></tr>
|
374 |
-
<tr>
|
375 |
-
<td><a href="https://huggingface.co/inceptionai/jais-family-6p7b-chat" target="_blank">jais-family-6p7b-chat</a></td>
|
376 |
-
<td>39.96</td>
|
377 |
-
<td>41.57</td>
|
378 |
-
<td>51.22</td>
|
379 |
-
</tr>
|
380 |
-
<tr>
|
381 |
-
<td><a href="https://huggingface.co/inceptionai/jais-adapted-7b-chat" target="_blank">jais-adapted-7b-chat</a></td>
|
382 |
-
<td>39.30</td>
|
383 |
-
<td>35.19</td>
|
384 |
-
<td>43.67</td>
|
385 |
-
</tr>
|
386 |
-
<tr>
|
387 |
-
<td><a href="https://huggingface.co/inceptionai/jais-family-13b-chat" target="_blank">jais-family-13b-chat</a></td>
|
388 |
-
<td>45.11</td>
|
389 |
-
<td>43.90</td>
|
390 |
-
<td>58.67</td>
|
391 |
-
</tr>
|
392 |
-
<tr>
|
393 |
-
<td><a href="https://huggingface.co/inceptionai/jais-adapted-13b-chat" target="_blank">jais-adapted-13b-chat</a></td>
|
394 |
-
<td>45.20</td>
|
395 |
-
<td>40.65</td>
|
396 |
-
<td>49.67</td>
|
397 |
-
</tr>
|
398 |
-
<tr>
|
399 |
-
<td><a href="https://huggingface.co/FreedomIntelligence/AceGPT-7B-chat" target="_blank">AceGPT-7b-chat</a></td>
|
400 |
-
<td>35.98</td>
|
401 |
-
<td>36.57</td>
|
402 |
-
<td>30.11</td>
|
403 |
-
</tr>
|
404 |
-
<tr>
|
405 |
-
<td><a href="https://huggingface.co/FreedomIntelligence/AceGPT-13B-chat" target="_blank">AceGPT-13b-chat</a></td>
|
406 |
-
<td>41.09</td>
|
407 |
-
<td>38.35</td>
|
408 |
-
<td>33.11</td>
|
409 |
-
</tr>
|
410 |
-
<tr>
|
411 |
-
<td><a href="https://huggingface.co/google/gemma-2-9b-it" target="_blank">gemma-2-9b-it</a></td>
|
412 |
-
<td>35.91</td>
|
413 |
-
<td>42.43</td>
|
414 |
-
<td>31.00</td>
|
415 |
-
</tr>
|
416 |
-
<tr>
|
417 |
-
<td><a href="meta-llama/Meta-Llama-3.1-8B-Instruct" target="_blank">Llama-3.1-8B-Instruct</a></td>
|
418 |
-
<td>44.13</td>
|
419 |
-
<td>38.24</td>
|
420 |
-
<td>47.00</td>
|
421 |
-
</tr>
|
422 |
-
<tr>
|
423 |
-
<td><strong><a href="https://huggingface.co/MBZUAI-Paris/Atlas-Chat-9B" target="_blank">Atlas-Chat-9B</a></strong></td>
|
424 |
-
<td><b>58.23</td>
|
425 |
-
<td><b>57.75</td>
|
426 |
-
<td><b>74.56</td>
|
427 |
-
</tr>
|
428 |
-
<tr style="border-top: 4px solid;"></tr>
|
429 |
-
<tr>
|
430 |
-
<td><a href="https://huggingface.co/inceptionai/jais-family-30b-8k-chat" target="_blank">jais-family-30b-8k-chat</a></td>
|
431 |
-
<td>51.88</td>
|
432 |
-
<td>35.61</td>
|
433 |
-
<td>65.67</td>
|
434 |
-
</tr>
|
435 |
-
<tr>
|
436 |
-
<td><a href="https://huggingface.co/google/gemma-2-27b-it" target="_blank">gemma-2-27b-it</a></td>
|
437 |
-
<td>36.47</td>
|
438 |
-
<td>37.04</td>
|
439 |
-
<td>35.78</td>
|
440 |
-
</tr>
|
441 |
-
<tr>
|
442 |
-
<td><strong><a href="https://huggingface.co/MBZUAI-Paris/Atlas-Chat-27B" target="_blank">Atlas-Chat-27B</a></strong></td>
|
443 |
-
<td><b>61.95</td>
|
444 |
-
<td><b>48.37</td>
|
445 |
-
<td><b>75.67</td>
|
446 |
-
</tr>
|
447 |
-
|
448 |
-
|
449 |
-
|
450 |
-
</table>
|
451 |
-
|
452 |
-
**Standard NLP Tasks:**
|
453 |
-
<table>
|
454 |
-
<tr>
|
455 |
-
<td rowspan="2">Model</td>
|
456 |
-
<td colspan="2"><a href="https://huggingface.co/datasets/MBZUAI-Paris/DarijaBench" target="_blank">DODa-10k (Translation)</a></td>
|
457 |
-
<td colspan="2"><a href="https://huggingface.co/datasets/MBZUAI-Paris/DarijaBench" target="_blank">MADAR (Translation)</a></td>
|
458 |
-
<td colspan="2"><a href="https://huggingface.co/datasets/MBZUAI-Paris/DarijaBench" target="_blank">FLORES+ (Translation)</a></td>
|
459 |
-
<td colspan="2"><a href="https://huggingface.co/datasets/MBZUAI-Paris/DarijaBench" target="_blank">NLLB-Seed (Translation)</a></td>
|
460 |
-
<td colspan="2"><a href="https://huggingface.co/datasets/MBZUAI-Paris/DarijaBench" target="_blank">DODa-10k (Transliteration)</a></td>
|
461 |
-
<td rowspan="2"><a href="https://huggingface.co/datasets/MBZUAI-Paris/DarijaBench" target="_blank">MArSum (Summarization)</a><br/>(LLM as a judge)</td>
|
462 |
-
<td rowspan="2"><a href="https://huggingface.co/datasets/MBZUAI-Paris/DarijaBench" target="_blank">Sentiment Analysis</a></td>
|
463 |
-
</tr>
|
464 |
-
<tr>
|
465 |
-
<td>BLEU</td>
|
466 |
-
<td>chrF</td>
|
467 |
-
<td>BLEU</td>
|
468 |
-
<td>chrF</td>
|
469 |
-
<td>BLEU</td>
|
470 |
-
<td>chrF</td>
|
471 |
-
<td>BLEU</td>
|
472 |
-
<td>chrF</td>
|
473 |
-
<td>BLEU</td>
|
474 |
-
<td>chrF</td>
|
475 |
-
</tr>
|
476 |
-
<tr>
|
477 |
-
<td><a href="https://huggingface.co/inceptionai/jais-family-1p3b-chat" target="_blank">jais-family-1p3b-chat</a></td>
|
478 |
-
<td>00.13</td>
|
479 |
-
<td>06.18</td>
|
480 |
-
<td>00.50</td>
|
481 |
-
<td>15.43</td>
|
482 |
-
<td>02.44</td>
|
483 |
-
<td>19.14</td>
|
484 |
-
<td>01.99</td>
|
485 |
-
<td>12.60</td>
|
486 |
-
<td>00.01</td>
|
487 |
-
<td>03.01</td>
|
488 |
-
<td>00.50</td>
|
489 |
-
<td>45.29</td>
|
490 |
-
</tr>
|
491 |
-
<tr>
|
492 |
-
<td><a href="https://huggingface.co/inceptionai/jais-family-2p7b-chat" target="_blank">jais-family-2p7b-chat</a></td>
|
493 |
-
<td>00.25</td>
|
494 |
-
<td>07.46</td>
|
495 |
-
<td>00.62</td>
|
496 |
-
<td>16.36</td>
|
497 |
-
<td>04.25</td>
|
498 |
-
<td>18.22</td>
|
499 |
-
<td>03.10</td>
|
500 |
-
<td>08.19</td>
|
501 |
-
<td>00.01</td>
|
502 |
-
<td>03.27</td>
|
503 |
-
<td>00.90</td>
|
504 |
-
<td>51.56</td>
|
505 |
-
</tr>
|
506 |
-
<tr>
|
507 |
-
<td><a href="https://huggingface.co/google/gemma-2-2b-it" target="_blank">gemma-2-2b-it</a></td>
|
508 |
-
<td>00.10</td>
|
509 |
-
<td>04.96</td>
|
510 |
-
<td>00.12</td>
|
511 |
-
<td>06.66</td>
|
512 |
-
<td>01.55</td>
|
513 |
-
<td>18.59</td>
|
514 |
-
<td>02.78</td>
|
515 |
-
<td>23.69</td>
|
516 |
-
<td>00.01</td>
|
517 |
-
<td>02.08</td>
|
518 |
-
<td>06.80</td>
|
519 |
-
<td>53.36</td>
|
520 |
-
</tr>
|
521 |
-
<tr>
|
522 |
-
<td><a href="meta-llama/Llama-3.2-1B-Instruct" target="_blank">Llama-3.2-1B-Instruct</a></td>
|
523 |
-
<td>00.07</td>
|
524 |
-
<td>05.95</td>
|
525 |
-
<td>00.80</td>
|
526 |
-
<td>18.71</td>
|
527 |
-
<td>04.53</td>
|
528 |
-
<td>18.39</td>
|
529 |
-
<td>04.52</td>
|
530 |
-
<td>17.06</td>
|
531 |
-
<td>00.02</td>
|
532 |
-
<td>03.74</td>
|
533 |
-
<td>08.23</td>
|
534 |
-
<td>46.27</td>
|
535 |
-
</tr>
|
536 |
-
<tr>
|
537 |
-
<td><a href="meta-llama/Llama-3.2-3B-Instruct" target="_blank">Llama-3.2-3B-Instruct</a></td>
|
538 |
-
<td>00.62</td>
|
539 |
-
<td>13.67</td>
|
540 |
-
<td>01.18</td>
|
541 |
-
<td>22.12</td>
|
542 |
-
<td>08.59</td>
|
543 |
-
<td>35.21</td>
|
544 |
-
<td>13.75</td>
|
545 |
-
<td>43.63</td>
|
546 |
-
<td>00.21</td>
|
547 |
-
<td>09.68</td>
|
548 |
-
<td>08.23</td>
|
549 |
-
<td>49.20</td>
|
550 |
-
</tr>
|
551 |
-
<tr>
|
552 |
-
<td><strong><a href="https://huggingface.co/MBZUAI-Paris/Atlas-Chat-2B" target="_blank">Atlas-Chat-2B</a></strong></td>
|
553 |
-
<td><b>22.76</td>
|
554 |
-
<td><b>44.86</td>
|
555 |
-
<td><b>16.67</td>
|
556 |
-
<td><b>41.64</td>
|
557 |
-
<td><b>14.92</td>
|
558 |
-
<td><b>43.03</td>
|
559 |
-
<td><b>23.88</td>
|
560 |
-
<td><b>52.19</td>
|
561 |
-
<td><b>08.18</td>
|
562 |
-
<td><b>21.54</td>
|
563 |
-
<td><b>55.22</td>
|
564 |
-
<td><b>73.99</td>
|
565 |
-
</tr>
|
566 |
-
<tr style="border-top: 4px solid;"></tr>
|
567 |
-
<tr>
|
568 |
-
<td><a href="https://huggingface.co/inceptionai/jais-family-6p7b-chat" target="_blank">jais-family-6p7b-chat</a></td>
|
569 |
-
<td>00.73</td>
|
570 |
-
<td>11.85</td>
|
571 |
-
<td>01.88</td>
|
572 |
-
<td>23.22</td>
|
573 |
-
<td>04.25</td>
|
574 |
-
<td>18.22</td>
|
575 |
-
<td>04.62</td>
|
576 |
-
<td>20.22</td>
|
577 |
-
<td>00.02</td>
|
578 |
-
<td>03.79</td>
|
579 |
-
<td>03.02</td>
|
580 |
-
<td>56.78</td>
|
581 |
-
</tr>
|
582 |
-
<tr>
|
583 |
-
<td><a href="https://huggingface.co/inceptionai/jais-adapted-7b-chat" target="_blank">jais-adapted-7b-chat</a></td>
|
584 |
-
<td>00.60</td>
|
585 |
-
<td>09.43</td>
|
586 |
-
<td>03.45</td>
|
587 |
-
<td>25.88</td>
|
588 |
-
<td>07.25</td>
|
589 |
-
<td>23.21</td>
|
590 |
-
<td>01.25</td>
|
591 |
-
<td>02.22</td>
|
592 |
-
<td>00.04</td>
|
593 |
-
<td>03.24</td>
|
594 |
-
<td>02.82</td>
|
595 |
-
<td>52.72</td>
|
596 |
-
</tr>
|
597 |
-
<tr>
|
598 |
-
<td><a href="https://huggingface.co/inceptionai/jais-family-13b-chat" target="_blank">jais-family-13b-chat</a></td>
|
599 |
-
<td>00.92</td>
|
600 |
-
<td>11.71</td>
|
601 |
-
<td>04.01</td>
|
602 |
-
<td>28.48</td>
|
603 |
-
<td>05.70</td>
|
604 |
-
<td>27.24</td>
|
605 |
-
<td>04.50</td>
|
606 |
-
<td>22.56</td>
|
607 |
-
<td>00.03</td>
|
608 |
-
<td>03.57</td>
|
609 |
-
<td>01.77</td>
|
610 |
-
<td>41.73</td>
|
611 |
-
</tr>
|
612 |
-
<tr>
|
613 |
-
<td><a href="https://huggingface.co/inceptionai/jais-adapted-13b-chat" target="_blank">jais-adapted-13b-chat</a></td>
|
614 |
-
<td>00.87</td>
|
615 |
-
<td>10.52</td>
|
616 |
-
<td>04.02</td>
|
617 |
-
<td>25.29</td>
|
618 |
-
<td>06.66</td>
|
619 |
-
<td>23.46</td>
|
620 |
-
<td>20.14</td>
|
621 |
-
<td>47.87</td>
|
622 |
-
<td>0.04</td>
|
623 |
-
<td>04.77</td>
|
624 |
-
<td>01.92</td>
|
625 |
-
<td>66.68</td>
|
626 |
-
</tr>
|
627 |
-
<tr>
|
628 |
-
<td><a href="https://huggingface.co/FreedomIntelligence/AceGPT-7B-chat" target="_blank">AceGPT-7b-chat</a></td>
|
629 |
-
<td>00.44</td>
|
630 |
-
<td>11.33</td>
|
631 |
-
<td>01.05</td>
|
632 |
-
<td>19.24</td>
|
633 |
-
<td>06.92</td>
|
634 |
-
<td>36.03</td>
|
635 |
-
<td>11.05</td>
|
636 |
-
<td>44.55</td>
|
637 |
-
<td>00.06</td>
|
638 |
-
<td>04.74</td>
|
639 |
-
<td>02.28</td>
|
640 |
-
<td>40.23</td>
|
641 |
-
</tr>
|
642 |
-
<tr>
|
643 |
-
<td><a href="https://huggingface.co/FreedomIntelligence/AceGPT-13B-chat" target="_blank">AceGPT-13b-chat</a></td>
|
644 |
-
<td>00.98</td>
|
645 |
-
<td>16.70</td>
|
646 |
-
<td>00.81</td>
|
647 |
-
<td>20.23</td>
|
648 |
-
<td>08.73</td>
|
649 |
-
<td>40.76</td>
|
650 |
-
<td>14.02</td>
|
651 |
-
<td>48.28</td>
|
652 |
-
<td>00.12</td>
|
653 |
-
<td>06.32</td>
|
654 |
-
<td>02.80</td>
|
655 |
-
<td>59.58</td>
|
656 |
-
</tr>
|
657 |
-
<tr>
|
658 |
-
<td><a href="https://huggingface.co/google/gemma-2-9b-it" target="_blank">gemma-2-9b-it</a></td>
|
659 |
-
<td>03.10</td>
|
660 |
-
<td>19.16</td>
|
661 |
-
<td>01.72</td>
|
662 |
-
<td>24.35</td>
|
663 |
-
<td>05.18</td>
|
664 |
-
<td>36.96</td>
|
665 |
-
<td>08.23</td>
|
666 |
-
<td>43.57</td>
|
667 |
-
<td>00.17</td>
|
668 |
-
<td>09.14</td>
|
669 |
-
<td>13.81</td>
|
670 |
-
<td>59.87</td>
|
671 |
-
</tr>
|
672 |
-
<tr>
|
673 |
-
<td><a href="meta-llama/Meta-Llama-3.1-8B-Instruct" target="_blank">Llama-3.1-8B-Instruct</a></td>
|
674 |
-
<td>00.92</td>
|
675 |
-
<td>14.19</td>
|
676 |
-
<td>01.46</td>
|
677 |
-
<td>23.82</td>
|
678 |
-
<td>08.89</td>
|
679 |
-
<td>33.08</td>
|
680 |
-
<td>11.85</td>
|
681 |
-
<td>35.51</td>
|
682 |
-
<td>00.11</td>
|
683 |
-
<td>06.02</td>
|
684 |
-
<td>01.28</td>
|
685 |
-
<td>44.08</td>
|
686 |
-
</tr>
|
687 |
-
<tr>
|
688 |
-
<td><strong><a href="https://huggingface.co/MBZUAI-Paris/Atlas-Chat-9B" target="_blank">Atlas-Chat-9B</a></strong></td>
|
689 |
-
<td><b>28.08</td>
|
690 |
-
<td><b>50.48</td>
|
691 |
-
<td><b>18.16</td>
|
692 |
-
<td><b>43.91</td>
|
693 |
-
<td><b>18.63</td>
|
694 |
-
<td><b>47.53</td>
|
695 |
-
<td><b>29.98</td>
|
696 |
-
<td><b>58.26</td>
|
697 |
-
<td><b>22.08</td>
|
698 |
-
<td><b>34.17</td>
|
699 |
-
<td><b>59.76</td>
|
700 |
-
<td><b>81.89</td>
|
701 |
-
</tr>
|
702 |
-
<tr style="border-top: 4px solid;"></tr>
|
703 |
-
<tr>
|
704 |
-
<td><a href="https://huggingface.co/inceptionai/jais-family-30b-8k-chat" target="_blank">jais-family-30b-8k-chat</a></td>
|
705 |
-
<td>01.10</td>
|
706 |
-
<td>14.40</td>
|
707 |
-
<td>01.67</td>
|
708 |
-
<td>23.37</td>
|
709 |
-
<td>08.52</td>
|
710 |
-
<td>35.41</td>
|
711 |
-
<td>13.71</td>
|
712 |
-
<td>41.33</td>
|
713 |
-
<td>00.05</td>
|
714 |
-
<td>04.48</td>
|
715 |
-
<td>00.46</td>
|
716 |
-
<td>56.73</td>
|
717 |
-
</tr>
|
718 |
-
<tr>
|
719 |
-
<td><a href="https://huggingface.co/google/gemma-2-27b-it" target="_blank">gemma-2-27b-it</a></td>
|
720 |
-
<td>00.67</td>
|
721 |
-
<td>13.04</td>
|
722 |
-
<td>01.74</td>
|
723 |
-
<td>24.63</td>
|
724 |
-
<td>05.17</td>
|
725 |
-
<td>37.08</td>
|
726 |
-
<td>07.36</td>
|
727 |
-
<td>42.49</td>
|
728 |
-
<td>00.03</td>
|
729 |
-
<td>04.94</td>
|
730 |
-
<td>11.10</td>
|
731 |
-
<td>57.59</td>
|
732 |
-
</tr>
|
733 |
-
<tr>
|
734 |
-
<td><strong><a href="https://huggingface.co/MBZUAI-Paris/Atlas-Chat-27B" target="_blank">Atlas-Chat-27B</a></strong></td>
|
735 |
-
<td><b>29.55</td>
|
736 |
-
<td><b>51.74</td>
|
737 |
-
<td><b>19.66</td>
|
738 |
-
<td><b>45.65</td>
|
739 |
-
<td><b>20.34</td>
|
740 |
-
<td><b>49.19</td>
|
741 |
-
<td><b>31.61</td>
|
742 |
-
<td><b>59.37</td>
|
743 |
-
<td><b>33.03</td>
|
744 |
-
<td><b>40.95</td>
|
745 |
-
<td><b>60.70</td>
|
746 |
-
<td>73.00</td>
|
747 |
-
</tr>
|
748 |
-
|
749 |
-
|
750 |
-
|
751 |
-
</table>
|
752 |
-
-->
|
753 |
-
|
754 |
## Evaluation
|
755 |
The Atlas-Chat models were evaluated on a comprehensive suite of tasks using various datasets and benchmarks to assess their performance across multiple dimensions. These included tasks such as:
|
756 |
|
@@ -777,14 +337,14 @@ The models were compared against a collection of existing open-source Arabic mod
|
|
777 |
<tr>
|
778 |
<td><a href="https://huggingface.co/inceptionai/jais-family-1p3b-chat" target="_blank">jais-family-1p3b-chat</a></td>
|
779 |
<td>35.39</td>
|
780 |
-
<td>
|
781 |
<td>38.33</td>
|
782 |
<td>35.56</td>
|
783 |
</tr>
|
784 |
<tr>
|
785 |
<td><a href="https://huggingface.co/inceptionai/jais-family-2p7b-chat" target="_blank">jais-family-2p7b-chat</a></td>
|
786 |
<td>37.44</td>
|
787 |
-
<td>
|
788 |
<td>44.11</td>
|
789 |
<td>52.97</td>
|
790 |
</tr>
|
@@ -812,7 +372,7 @@ The models were compared against a collection of existing open-source Arabic mod
|
|
812 |
<tr>
|
813 |
<td><strong><a href="https://huggingface.co/MBZUAI-Paris/Atlas-Chat-2B" target="_blank">Atlas-Chat-2B</a></strong></td>
|
814 |
<td><b>44.97</b></td>
|
815 |
-
<td><b>
|
816 |
<td><b>53.89</b></td>
|
817 |
<td><b>92.31</b></td>
|
818 |
</tr>
|
@@ -820,35 +380,35 @@ The models were compared against a collection of existing open-source Arabic mod
|
|
820 |
<tr>
|
821 |
<td><a href="https://huggingface.co/inceptionai/jais-family-6p7b-chat" target="_blank">jais-family-6p7b-chat</a></td>
|
822 |
<td>39.96</td>
|
823 |
-
<td>
|
824 |
<td>51.22</td>
|
825 |
<td>65.18</td>
|
826 |
</tr>
|
827 |
<tr>
|
828 |
<td><a href="https://huggingface.co/inceptionai/jais-adapted-7b-chat" target="_blank">jais-adapted-7b-chat</a></td>
|
829 |
<td>39.30</td>
|
830 |
-
<td>
|
831 |
<td>43.67</td>
|
832 |
<td>61.84</td>
|
833 |
</tr>
|
834 |
<tr>
|
835 |
<td><a href="https://huggingface.co/inceptionai/jais-family-13b-chat" target="_blank">jais-family-13b-chat</a></td>
|
836 |
<td>45.11</td>
|
837 |
-
<td>
|
838 |
<td>58.67</td>
|
839 |
<td>69.93</td>
|
840 |
</tr>
|
841 |
<tr>
|
842 |
<td><a href="https://huggingface.co/inceptionai/jais-adapted-13b-chat" target="_blank">jais-adapted-13b-chat</a></td>
|
843 |
<td>45.20</td>
|
844 |
-
<td>
|
845 |
<td>49.67</td>
|
846 |
<td>77.52</td>
|
847 |
</tr>
|
848 |
<tr>
|
849 |
<td><a href="https://huggingface.co/FreedomIntelligence/AceGPT-7B-chat" target="_blank">AceGPT-7b-chat</a></td>
|
850 |
<td>35.98</td>
|
851 |
-
<td>
|
852 |
<td>30.11</td>
|
853 |
<td>47.31</td>
|
854 |
</tr>
|
@@ -862,21 +422,21 @@ The models were compared against a collection of existing open-source Arabic mod
|
|
862 |
<tr>
|
863 |
<td><a href="https://huggingface.co/google/gemma-2-9b-it" target="_blank">gemma-2-9b-it</a></td>
|
864 |
<td>35.91</td>
|
865 |
-
<td>
|
866 |
<td>31.00</td>
|
867 |
<td>90.86</td>
|
868 |
</tr>
|
869 |
<tr>
|
870 |
<td><a href="meta-llama/Meta-Llama-3.1-8B-Instruct" target="_blank">Llama-3.1-8B-Instruct</a></td>
|
871 |
<td>44.13</td>
|
872 |
-
<td>
|
873 |
<td>47.00</td>
|
874 |
<td>78.08</td>
|
875 |
</tr>
|
876 |
<tr>
|
877 |
<td><strong><a href="https://huggingface.co/MBZUAI-Paris/Atlas-Chat-9B" target="_blank">Atlas-Chat-9B</a></strong></td>
|
878 |
<td><b>58.23</b></td>
|
879 |
-
<td><b>
|
880 |
<td><b>74.56</b></td>
|
881 |
<td><b>95.62</b></td>
|
882 |
</tr>
|
|
|
311 |
Atlas-Chat models are based on Gemma 2 models. The Atlas-Chat models were trained using 8 Nvidia's A100 80 GB GPUs in parallel using FSDP on AWS Sagemaker. The model is trained using HuggingFace transformers and parameter-efficient fine-tuning with LoRA rank of 256.
|
312 |
|
313 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
314 |
## Evaluation
|
315 |
The Atlas-Chat models were evaluated on a comprehensive suite of tasks using various datasets and benchmarks to assess their performance across multiple dimensions. These included tasks such as:
|
316 |
|
|
|
337 |
<tr>
|
338 |
<td><a href="https://huggingface.co/inceptionai/jais-family-1p3b-chat" target="_blank">jais-family-1p3b-chat</a></td>
|
339 |
<td>35.39</td>
|
340 |
+
<td>27.71</td>
|
341 |
<td>38.33</td>
|
342 |
<td>35.56</td>
|
343 |
</tr>
|
344 |
<tr>
|
345 |
<td><a href="https://huggingface.co/inceptionai/jais-family-2p7b-chat" target="_blank">jais-family-2p7b-chat</a></td>
|
346 |
<td>37.44</td>
|
347 |
+
<td>29.10</td>
|
348 |
<td>44.11</td>
|
349 |
<td>52.97</td>
|
350 |
</tr>
|
|
|
372 |
<tr>
|
373 |
<td><strong><a href="https://huggingface.co/MBZUAI-Paris/Atlas-Chat-2B" target="_blank">Atlas-Chat-2B</a></strong></td>
|
374 |
<td><b>44.97</b></td>
|
375 |
+
<td><b>35.08</b></td>
|
376 |
<td><b>53.89</b></td>
|
377 |
<td><b>92.31</b></td>
|
378 |
</tr>
|
|
|
380 |
<tr>
|
381 |
<td><a href="https://huggingface.co/inceptionai/jais-family-6p7b-chat" target="_blank">jais-family-6p7b-chat</a></td>
|
382 |
<td>39.96</td>
|
383 |
+
<td>32.64</td>
|
384 |
<td>51.22</td>
|
385 |
<td>65.18</td>
|
386 |
</tr>
|
387 |
<tr>
|
388 |
<td><a href="https://huggingface.co/inceptionai/jais-adapted-7b-chat" target="_blank">jais-adapted-7b-chat</a></td>
|
389 |
<td>39.30</td>
|
390 |
+
<td>29.55</td>
|
391 |
<td>43.67</td>
|
392 |
<td>61.84</td>
|
393 |
</tr>
|
394 |
<tr>
|
395 |
<td><a href="https://huggingface.co/inceptionai/jais-family-13b-chat" target="_blank">jais-family-13b-chat</a></td>
|
396 |
<td>45.11</td>
|
397 |
+
<td>33.98</td>
|
398 |
<td>58.67</td>
|
399 |
<td>69.93</td>
|
400 |
</tr>
|
401 |
<tr>
|
402 |
<td><a href="https://huggingface.co/inceptionai/jais-adapted-13b-chat" target="_blank">jais-adapted-13b-chat</a></td>
|
403 |
<td>45.20</td>
|
404 |
+
<td>32.84</td>
|
405 |
<td>49.67</td>
|
406 |
<td>77.52</td>
|
407 |
</tr>
|
408 |
<tr>
|
409 |
<td><a href="https://huggingface.co/FreedomIntelligence/AceGPT-7B-chat" target="_blank">AceGPT-7b-chat</a></td>
|
410 |
<td>35.98</td>
|
411 |
+
<td>30.33</td>
|
412 |
<td>30.11</td>
|
413 |
<td>47.31</td>
|
414 |
</tr>
|
|
|
422 |
<tr>
|
423 |
<td><a href="https://huggingface.co/google/gemma-2-9b-it" target="_blank">gemma-2-9b-it</a></td>
|
424 |
<td>35.91</td>
|
425 |
+
<td>32.19</td>
|
426 |
<td>31.00</td>
|
427 |
<td>90.86</td>
|
428 |
</tr>
|
429 |
<tr>
|
430 |
<td><a href="meta-llama/Meta-Llama-3.1-8B-Instruct" target="_blank">Llama-3.1-8B-Instruct</a></td>
|
431 |
<td>44.13</td>
|
432 |
+
<td>31.40</td>
|
433 |
<td>47.00</td>
|
434 |
<td>78.08</td>
|
435 |
</tr>
|
436 |
<tr>
|
437 |
<td><strong><a href="https://huggingface.co/MBZUAI-Paris/Atlas-Chat-9B" target="_blank">Atlas-Chat-9B</a></strong></td>
|
438 |
<td><b>58.23</b></td>
|
439 |
+
<td><b>43.65</b></td>
|
440 |
<td><b>74.56</b></td>
|
441 |
<td><b>95.62</b></td>
|
442 |
</tr>
|