walterpcasas/fatima-test

Fatima Fellowship - Dataset for Model Evaluation

Model: Tucano2-0.6B-Base

The model evaluated is Tucano2-0.6B-Base, a decoder-only transformer pretrained primarily on Portuguese and English as part of the Polygl0t initiative.
It is a base model, meaning it was trained for general language modeling and not instruction-tuned for specific tasks.

How the model was loaded

The model was loaded and tested in a Google Colab notebook using Hugging Face Transformers.

from transformers import GenerationConfig, TextGenerationPipeline, AutoTokenizer, AutoModelForCausalLM
import torch
from datasets import load_dataset

model_id = "Polygl0t/Tucano2-0.6B-Base"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

prompts = dataset['train']['input']

generation_config = GenerationConfig(
    **{
    "do_sample": True,
    "max_new_tokens": 150,
    "renormalize_logits": True,
    "repetition_penalty": 1.2,
    "temperature": 0.1,
    "top_k": 50,
    "top_p": 1.0,
    "use_cache": True,
  }
)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
generator = TextGenerationPipeline(model=model, task="text-generation", tokenizer=tokenizer, device=device)

outputs = generator(list(prompts), generation_config=generation_config)

model_outputs = []

for prompt, output in zip(prompts, outputs):
    full_text = output[0]["generated_text"]
    completion = full_text[len(prompt):].strip()
    model_outputs.append(completion)

dataset["train"] = dataset["train"].add_column("model_result", model_outputs)

dataset["train"].to_pandas()

Observed blind spots

Based on the 10 evaluation examples, several weaknesses appeared:

1. Poor instruction following: The model frequently ignores strict constraints such as “one word”, “number only”, or “YES/NO”.

2. Overly verbose responses: Even when a short answer is requested, the model tends to generate explanations.

3. Prompt drift: Some outputs are unrelated to the prompt, indicating weak task grounding.

4. Failures on deterministic tasks: The model struggles with tasks requiring exact outputs, such as copying text, counting characters, or selecting a single category.

5. Literal translation errors: Some translations transfer Portuguese structure directly into English (e.g., “I have 25 years”).

6. Quiz-style bias: Several outputs generate multiple-choice formats (A/B/C/D), suggesting training data containing educational material.

How the model could be improved

The model would benefit from instruction tuning, especially on datasets that emphasize:

• strict output formatting
• short answers
• translation pairs (Portuguese ↔ English)
• classification and reasoning tasks

Potential dataset sources include:

• instruction-following datasets
• translation corpora
• QA datasets with constrained answers

How to assemble or find such a dataset

A suitable dataset could be created by combining existing public datasets and manually written examples. For example, Portuguese–English translation datasets and instruction-following datasets can be found on platforms like Hugging Face. Additional examples could be created by writing prompts that require strict answers (such as one word, a number, or YES/NO) and adding the correct outputs. These examples could then be reviewed and combined into a dataset that helps the model learn to follow instructions and produce short, accurate responses. It is also possible to generate more examples using a larger commercial LLM and then review and filter the generated data before adding it to the dataset.

Estimated dataset size

For a model of this size (0.6B parameters), approximately 100k high-quality examples would likely be sufficient.