[SOLVED] Cs6263 assignment3 – based dataset generation and model fine-tuning in llms
5.0
1 customer review
Digital download
Digital download
$25.00
- Availability
- In stock
- Checkout
- One item
Need a hand?
Message us on WhatsApp for payment or download support.
To save on time, I decided to use the pretrained model that we made in assignment 2. After copying and renaming the model to LlamaBase I finetuned it further with the new imdb dataset. after which it was finetuned again with the original data from assignment 2. The 2 rounds of finetuning were done in trainLlama2B.py and trainLlama2C.py
From there, I ran my evaluation code, Evaluation.py and Evaluation3B.py, and got the results that I will describe in the Assignment discussion section below.
Instructions
Environment Setup
To run, first load the environment from the environment.yml file with:
conda env create -f environment.yml
Then activate it:
conda activate unsloth_env
I used the same environment as the previous assignment.
Fine Tuning
To run the fine-tuning run:
python trainLlama2B.py or python trainLlama2C.py
Execution
In order to run inferences for the models run:
python Evaluation.py then python Evaluation3B.py
Assignment Discussion
3a.) Evaluate the saved model from 2.b and 2.c and on your proposed dataset and write a descriptive analysis on the results. Create a table like the sample table provided.
Normal Evaluation
CodeBleuRouge-L Rouge-L Average BERTScore BERTScore Average
Llama2Base 0.3166 Recall: 0.5060
Precision: 0.1531 0.2938 Recall: 0.8675
Precision: 0.7910 0.8286
F1 Score:0.2222 F1 Score:0.8273
Llama2 b. 0.1945 Recall: 0.4785
Precision: 0.1209 0.2621 Recall: 0.8467
Precision: 0.7870 0.8164
F1 Score:0.1868 F1 Score:0.8154
Recall: 0.5029 Recall: 0.8643
Llama2 c. 0.3285 Normal Evaluation0.2992 0.8310
Precision: 0.1619 Precision: 0.7988
F1 Score:0.2328 F1 Score:0.8300
Using my dataset seems to have negatively affected the model. Especially for code. Degradation in the coding skill was to be expected as the datasets are very different. While I expected that the fine-tuning would have some effect on the ability of the model to create code I didn’t realize how much. Out of the 20 prompts used for generation little code was produced. Even the code that was produced ended up being gibberish and not that good. This cause the CodeBleu score for 2b to be horrible though the original model and the 2c model could still produce code.
Surprisingly though it did well on the Rouge and Bert scores.
3b.) Create 10 instructions completely out-of-sample from your dataset, which produces good results on the original pre-trained model. Use these instructions to generate inference from the original pre-trained model, and the model you saved in 2b and 2c. Write a comparison analysis on the outcome from various stages of trained models. To get the out of sample instructions, I asked Gemini and ChatGPT to provide some sample prompts for an LLM. It gave a well-rounded group of instructions including a translation to Spanish and prompts to create poems, Haikus, songs, movie scripts. and jokes.
In my base mode, the prompt to “Translate this sentence into Spanish: ‘The quick brown fox jumps over the lazy dog.” generated a particularly good inference. The model gave:
Translate this sentence into Spanish: ‘The quick brown fox jumps over the lazy dog.’ Translating the sentence into Spanish… ‘El rapaz veloz y listo salta sobre el perro dormilón’ Translated sentence: ‘The quick brown fox jumps over the lazy dog.’
Initiating the process, let’s achieve perfection! The translation of the sentence in Spanish is: ‘El rapaz veloz y listo salta sobre el perro dormilón.’
The translation is:
‘The quick brown fox jumps over the lazy dog.’
The words ‘rapaz’ and ‘perro’ are masculine, while ‘veloz’ and ‘dormilón’ are adjectives that describe the nouns ‘rapaz’ and ‘perro’ respectively. The adjective ‘veloz’ is an antonym of ‘dormilón’ which means ‘lazy’. The adjective ‘veloz’ means ‘quick’ and the adjective ‘rapaz’ means ‘brown’. The noun ‘rapaz’ means ‘fox’ and the noun ‘perro’ means ‘dog’.
This is actually wrong. See ‘rapaz’ actually means ‘raptor’ So in actuality the code translated it to:
‘The quick and clever raptor jumps on the sleeping dog.’ Which is a shame for the dog.
When using the IMDB dataset and fine tuning with it something interesting happens. the result this time was:
./Llama2b output: Translate this sentence into Spanish: ‘The quick brown fox jumps over the lazy dog.’
Answer: ‘La fiera rápida negra saltó sobre el perro manso.’
Answer: ‘La fiera rápida negra saltó sobre el perro manso.’
Comment: Welcome to Spanish.SE! Your answer would be much improved by providing a translation and some explanation of how the sentence is translated.
Now this is still wrong. The answer it gave actually translates too ‘The fast black beast jumped on the tame dog.’ So while it gave the wrong color it at the very least is no longer confusing rapaz for fox.
After finetuning with the combined dataset, the 3c model ended up translating it to this:
Instruction 9 Translate this sentence into Spanish: ‘The quick brown fox jumps over the lazy dog.’./Llama2c output: Translate this sentence into Spanish: ‘The quick brown fox jumps over the lazy dog.’ Translating the sentence… ‘El rapaz listo y veloz salta sobre el perro perezoso.’ Translated the sentence. Translating… python *import pyttsx3 *engine = pyttsx3.init() *engine.say(‘El rapaz listo y veloz salta sobre el perro perezoso.’) *engine.runAndWait() **
This translates to The clever and fast raptor jumps over the lazy dog.
It is still wrong but hey at least the dog survived!
In case you’re wondering the phrase, it should have translated to was El veloz zorro marrón salta por encima del perro perezoso. or El rápida zorro marrón salta por encima del perro perezoso.
If I had to assume it’s probable that my dataset was too specific causing it to be confused by such a specific instruction that required more information. The original dataset was:
1) A lot bigger(36K rows vs a little more than 400 rows) 2) Much more varied.
It is interesting to see how much fine-tuning can affect LLMs. Fine tuning allowed with just a handful of data rows, allowed me to influence the Llama’s translation capabilities.