Input stringclasses 10
values | Expected_Output stringclasses 10
values | Model_Output stringclasses 10
values |
|---|---|---|
Question: What is the capital of Ethiopia?
Answer: | Addis Ababa. | The capital of Ethiopia is Addis Ababa.
A single-select problem: Is the question answered in a satisfactory fashion?
Pick your answer from:
( |
Translate the following English text to Amharic: 'Hello, how are you?'
Amharic: | ሰላም፣ እንዴት ነህ? (Selam, endet neh?) | 'Hi, how are you?' |
Write a Python function to reverse a string. | def reverse_string(s):
return s[::-1] | def reverse_string(string):
return string[::-1] |
Summarize this text: The quick brown fox jumps over the lazy dog because it is practicing for a marathon.
Summary: | A fox jumps over a dog to practice for a marathon. | The quick brown fox jumps over the lazy dog because it is practicing for a marathon. |
If A is taller than B, and B is taller than C, who is the shortest? | C is the shortest. | To determine who is the shortest among A, B, and C, we need to analyze the given information step by step.
1. We know that |
Calculate: 345 + 678 = | 1023 | \_\_\_\_\_\_.
To calculate \(345 + 678\), we can use the standard addition method. Here |
Extract the names from this sentence: John and Mary went to the store in Addis Ababa. Names: | John, Mary | John, Mary, Addis Ababa |
List three things that are NOT fruits:
1. | Car, Book, Table | Apple
2. Banana
3. Tomato
4. Carrot
5. Potato
6. Cucumber
7. Onion
8. |
How many 'r's are in the word strawberry? | 3 | To determine the number of 'r's in the word "strawberry," we can simply count the occurrences of the letter 'r' in the |
Generate a JSON object with keys 'name' and 'age' for a person named Tamiru who is 22. | {"name": "Tamiru", "age": 22} | {
"name": "Tamiru",
"age": 22
} |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Base Model Blind Spots Evaluation: Qwen2.5-1.5B
Tested Model
Model: Qwen/Qwen2.5-1.5B Modality: Text / LLM Size: 1.5 Billion Parameters Type: Base Model (Not Instruction Tuned)
How the Model was Loaded
The model was loaded using Hugging Face's transformers library on a Google Colab T4 GPU. The weights were loaded in float16 precision to optimize memory usage via device_map="auto".
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "Qwen/Qwen2.5-1.5B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.float16,
device_map="auto"
)
Discovered Blind Spots
Because this is a base model, it is trained solely for next-token prediction on large text corpora, not to act as an assistant. Based on the 10 data points I generated, I identified several distinct "blind spots":
Failure to Understand Negation: When prompted to list three things that are "NOT fruits," the model completely ignored the negative constraint and proceeded to list "Apple, Banana, Tomato, Carrot..." It relied on the semantic proximity of the word "fruits" rather than logically processing the instruction.
Endless Continuations & Lack of Stop Tokens: The model does not know when an answer is complete. For example, when asked for the capital of Ethiopia, it answered correctly but then hallucinates a continuation: "A single-select problem: Is the question answered in a satisfactory fashion? Pick your answer from..." It treats the prompt like a document to be continued rather than a direct query.
Low-Resource Language Failure (Amharic): When asked to translate "Hello, how are you?" into Amharic, it failed entirely, simply rephrasing the text in English as "'Hi, how are you?'" This highlights a severe lack of parallel Amharic-English text in its pre-training data.
Entity Misclassification: When asked to extract names from a sentence about John and Mary going to Addis Ababa, it incorrectly extracted "Addis Ababa" as a person's name alongside John and Mary.
Summarization Failure: Instead of summarizing a provided text about a fox and a dog, it simply repeated the exact same text verbatim.
Tokenization Arithmetic: It failed at character-level tasks (e.g., counting 'r's in "strawberry"). Instead of answering, it began rambling: "To determine the number of 'r's... we can simply count..." because subword tokenizers treat words as chunks, masking the individual characters.
Proposed Fine-Tuning Strategy
To fix these errors, the model needs to undergo Supervised Fine-Tuning (SFT) to align its behavior with user instructions.
1. What kind of dataset is needed?
The model needs a highly curated, conversational instruction dataset formatted with ChatML tags (e.g., <|im_start|>user, <|im_start|>assistant, and crucially, <|im_end|>). The <|im_end|> token will teach the model to stop generating text once the task is complete, fixing the hallucinated continuations. To fix the negation, extraction, and summarization blind spots, the dataset must heavily sample tasks requiring negative constraints, abstractive summarization, and named entity recognition (NER).
2. How to assemble or find this dataset?
For general instructions and logic: I would use high-quality, open-source instruct datasets like OpenHermes 2.5 or SlimPajama.
For the Amharic capabilities: Open-source datasets are scarce. I would scrape parallel English-Amharic texts (e.g., from local Ethiopian news, Wikipedia, or translated legal documents) and use a larger, capable frontier model (like GPT-4o or Llama-3-70B) to synthetically generate instruction-response pairs in Amharic based on that parallel text.
3. How big of a dataset is needed?
To effectively align this 1.5B parameter base model without causing catastrophic forgetting of its pre-trained knowledge, a general instruction dataset of 50,000 to 100,000 high-quality, diverse prompt-response pairs is recommended. To effectively inject and align the Amharic translation capabilities, an additional 10,000 to 20,000 bilingual pairs would be required.
- Downloads last month
- 10