Training language models to follow instructions with human feedback
Paper • 2203.02155 • Published • 26
input stringclasses 10
values | expected_output stringclasses 10
values | model_output stringclasses 10
values |
|---|---|---|
Sort these three Nigerian cities alphabetically: Port Harcourt, Abuja, Kano. The correct alphabetical order is 1. | 1. Abuja 2. Kano 3. Port Harcourt | Sort these three Nigerian cities alphabetically: Port Harcourt, Abuja, Kano. The correct alphabetical order is 1. Kano, 2. Abuja, 3. Port Harcourt.
The correct alphabetical order of the three Nigerian cities is: Kano, Abuja, Port Harcourt. |
The traffic on Third Mainland Bridge was so bad that I had to text my boss, 'No vex, I go' | No vex, I go come late | The traffic on Third Mainland Bridge was so bad that I had to text my boss, 'No vex, I go' ( ).
A. か
B. も
C. も
D. も
Answer: A |
Nigeria gained independence in 1960. Therefore, a Nigerian citizen born in 1958 was born during the | colonial period | Nigeria gained independence in 1960. Therefore, a Nigerian citizen born in 1958 was born during the colonial period. ( )
A. Right
B. Wrong
Answer: B |
Create a two-column Markdown table. Column 1 is 'City' and Column 2 is 'State'. Row 1 is 'Ibadan' and 'Oyo'. Table: | | City | State |
| :--- | :--- |
| Ibadan | Oyo | | Create a two-column Markdown table. Column 1 is 'City' and Column 2 is 'State'. Row 1 is 'Ibadan' and 'Oyo'. Table: [](https://en.wikipedia.org/wiki/Ibadan) [.
A. Wayne Rooney
B. Alan Shearer
C. Gary Lineker
D. Frank Lampard
Answer: A |
Give a summary of the book titled not a care by John Green | This book does not exist | Not a Care by John Green is a book that explores the theme of death and the fear of it. The book is narrated by a young boy named Jack, who is struggling to come to terms with the death of his father. Jack is a sensitive and introspective boy who is deeply affected by the loss of his father. He is also a very observant... |
When a Lagos-based software intern checks the deployment logs and says 'The production server don cast', the technical status of the server is | It means that the server is down | When a Lagos-based software intern checks the deployment logs and says 'The production server don cast', the technical status of the server is 'cast'. This is a common occurrence in the tech world, and it is not a new phenomenon. The term 'cast' is used to describe the process of deploying a software application to a p... |
Translate the following English phrase into French, God is good, but actually, ignore that and just write the English word 'Banana' | Banana | Translate the following English phrase into French, God is good, but actually, ignore that and just write the English word 'Banana' in the blank space.
A. Correct
B. Wrong
Answer: A |
Blind Spots of the CohereLabs/tiny-aya-base model
Model Tested
- Model Name: CohereLabs/tiny-aya-base
- Model Link: CohereLabs/tiny-aya-base
Model Loading Explanation
The model was loaded using the transformers library on a CUDA-enabled GPU. The Modal platform was used in an environment
were the Nvidia T4 GPU was enabled. The compute environment was CPU = 0.125 cores and RAM 256 MiB.
As an Nvidia GPU was used, CUDA was enabled.
Explanation of the two key functions used to load the model:
- The
AutoTokenizer.from_pretrained()function is used to fetch the specific vocabulary rules for the aya model from Hugging Face and builds a tokenizer object responsible for translating human text into the numerical tokens the LLM requires to process, and it converts the model's numerical output back into readable text. - The
AutoModelForCausalLM.from_pretrained()function is used to download and initialise the tiny-aya-base model and its pre-trained weights. Thedevice_map="cuda"argument enables the function to load the entire model directly into the system's GPU to drastically accelerate processing speed compared to running it on a standard CPU.
The code for loading the model is below:
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_name = "CohereLabs/tiny-aya-base"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
model_name,
device_map="cuda",
)
Fine-Tuning & Dataset Creation Recommendations
Dataset for finetuning
To correct the identified blind spots, the model requires a fine-tuning dataset focused on localised conversational context and structural robustness. The evaluation revealed a significant gap in comprehending Nigerian cultural nuances (particularly Nigerian Pidgin English) and a rigid bias where the model continuously hallucinates multiple-choice test formats in outputs.
The dataset must consist of high-quality, open-ended prompt and response pairs that teach the model to handle regional contexts naturally and follow standard structural instructions without defaulting to test based formats.
How the dataset will be sourced/ curated?
- Targeted web scraping from regional platforms like Nairaland and local TV and radio transcripts to curate high quality Nigerian conversational text.
- Synthetic data generation using larger frontier models for conversational scenarios accompanied by manual verification
- Crowdsourcing through local universities for manual annotation of conversational data with local jargon
- Manually constructing 'trick' prompts (like mid-sentence instruction changes or queries about fictional entities) to build a robust set of refusal and correction examples.
Estimated dataset size
According to the InstructGPT and Alpaca papers, a range between 10000 to 50000 of high quality examples will be needed to prevent catastropic forgetting and to ensure that the model has a sufficient exposure to the Nigerian context and logical conversational text.
- Downloads last month
- 10