[meta] Retrain older high resource models

[eu9ene]

A meta-issue to track retraining of the older models like Italian, Portuguese, French, German, Spanish etc.

We'll need to retrain them to incorporate the latest robustness fixes. Also open-source datasets have improved since we trained them.

Those are high-resource languages with massive amounts of available training data, so training them is a bit more expensive than a typical language from our last big training. We'll need to do the optimizations first and also fix as many quality issues as we can before retraining not to do this again in the near future.

Must have

Preview Give feedback

1. MTData fails to unpack some datasets #942
   bug +1
2. Add bcp 47 code support in mtdata importer. #76
   bug data sources +2
3. Improve translation of short sentences #215
   quality +1
4. Use a more precise language identification model to clean datasets #248
   quality +1
5. Check if issues with short sentences were caused by bicleaner hard rules #903
   quality +1
6. Ensure presence or absence of periods doesn't trigger wrong translations #177
   quality +1
7. Single-side deduplication #928
   quality +1
8. Add aug-mix to mtdata devtest datasets #651
   enhancement good first issue +2
9. Figure out sacrebleu / dataset strategy #634
   enhancement +1
10. Corpora exclusion rules #940
    language-coverage +1
11. bicleaner-ai-classify intermittently fails to download fasttext model #631
    bug taskcluster +2
12. Rerun only tasks downstream from start-stage #1025
    taskcluster +1
13. Can't restart the pipeline to run distillation from where it left #711
    taskcluster +1
14. train-student OSError: [Errno 28] No space left on device #774
    bug +1
15. Add a cleaning rule for URL names, such as Amazon.com -> Amazon.it #758
    quality +1
16. Improve translation of URLs #736
    quality +1
17. Ensure numbers associated to units of measure are not translated as the same number with another unit of measure #577
    quality +1
18. Currency translation for English to German is incorrect #870
    feedback quality +2
19. Make sure square brackets instead of parentheses don't lead to wrong translations #800
    quality +1
20. Investigate how max_length affects training and adjust it #1038
    bug quality +2
21. dataset-opus-NLLB_v1-de-en fails #1040
    bug +1

Nice to have

Preview Give feedback

1. Teacher training configuration improvements #987
   quality +1
2. More corpora specific fixes #910
   quality +1
3. Translation settings for backtranslation are suboptimal #92
   bug quality +2
4. max-length and max-length-crop considered harmful #89
   bug quality +2
5. Reduce the number of vCPUs for bicleaner machines #394
   cost & perf +1
6. Full audit of resource utilization and adjustments #414
   cost & perf +1
7. Improve GPU utilization in student training #783
   cost & perf +1
8. Limit the amount of data used for distillation #905
   cost & perf +1
9. Do not run bicleaner step if the threshold is 0 #415
   cost & perf +1
10. start_stage often reruns amost all "evaluate" tasks #728
    cost & perf taskcluster +2
11. automatically upload important artifacts to a GCP bucket #466
    taskcluster +1
    
12. Improve translation of social posts #737
    quality +1
13. In the cleaning task, output statistics about filtered out/kept sentences and maybe attach list of filtered out sentences as an artifact #274
    enhancement quality +2

[marco-c]

After this is done, we'll have to verify the fix with #757 (comment).