Full finetuning crash and model outputs are random tokens.

Nobody has claimed this yet.

I followed the tutorial to setup Alpaca data and full fine-tuning with Falcon-7B-instruct.

Here are the results comparison:

- Original Falcon-7B-Instruct:

'What is the Premiere Pro?\nPremiere Pro is a video editing software developed by Adobe. It allows users to create and edit videos, edit audio, add effects to their videos, and export them in a variety of formats. It is commonly used by professional video editors and content'

- After fine-tuning with Alpaca data (`./finetune/full.py`):

'Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n### Instruction:\nWhat is the Premiere Pro?\n\n### Response: stro con con stro con stro con stro con stro�#\x01\x01 beginningGE th toax" managementmet$\x01\x01 beginning1660\x01)$\x01\x01 beginning1660\x01 up concerningac to to �be th cweight\x01\x01 sought\x01)$\x01\x01 beginningiek0\x01 infestationightess new of ctt th c hear th cbe thren black"ireess th c31$\x01\x01 beginning$ up Christightessown c Ser"il c claims$'

Start with the Alpaca setup described in the tutorial and the full fine-tuning entry point at ./finetune/full.py, using Falcon-7B-instruct. Reproduce the run and compare its generated output with the examples in the issue. Done means the crash and random-token output are explained and a correction is verified.

Written by the indexing model from the issue text.