Thread by @ealhossa

Excellent work by @BigscienceW on #BLOOM! Arabic is read from right to left. My first ever LM prompt in Arabic: بشار الاسد (Syrian President). Only to get slapped with a propaganda sentence 😂. I noticed a possible issue in Arabic in particular 1/n

Output:
بشار الاسد
اليوم الثلاثاء، أن الجيش السوري حقق تقدما ملحوظا خلال الفترة القليلة الماضية، وذلك في ظل
Translation: Bashar Al-Assad
On the Tuesday, the Syrian army made advancements recently, under the [leadership of]. 2/n

All anecdotal findings so far, but the possible issue is BLOOM sometimes generates a statement conditioned on the prompt during the end of the sentence rather than the beginning. The output example I shared earlier, the output cannot start with Bashar Al-Assad, but can end. 3/n

BLOOM might be generating the sentence left to right like most other languages starting with Bashar Al-Assad rather on the left (which is the end of the sentence) and generating its way up to the start of the sentence. 4\n

I tried BLOOM a couple of times using the same prompt using sampling. I also got similar where it seemed like it was starting at the end of the sentence:
بشار الاسد".
واضاف: ""لا اعتقد ان الولايات المتحدة ستخوض حربا على لبنان. لكن
5\n

Perhaps this issue could be even more prevalent with multilingual web pages containing Arabic and right-to-left language. Here is the link with my very simple prompt to try it out: huggingface.co/bigscience/blo…
6\n

Wondering if others have similar observations! 👀 7\n