rafio1020/qwen3-0.8b-base-blindspots

Qwen3.5-0.8B-Base Blind Spots Dataset

Model Tested

Qwen/Qwen3.5-0.8B-Base

• Parameters: 0.8B
• Type: Base language model (text completion, not instruction-tuned)
• Release: 2025

---

How I Loaded the Model

Tested on Kaggle using a T4 GPU:

!pip install -U git+https://github.com/huggingface/transformers accelerate bitsandbytes

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "Qwen/Qwen3.5-0.8B-Base"

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_id)

def test_blind_spot(prompt, max_new_tokens=100):
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    with torch.no_grad():
        output = model.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)
    decoded = tokenizer.decode(output[0], skip_special_tokens=True)
    return decoded[len(prompt):]  # Return only the newly generated part

Prompts were written as open-ended completions (e.g. "The capital of Australia is ") since this is a base model, not an instruction-following model.

---

Blind Spots Found

Testing revealed 3 major recurring failure patterns:

1. Format hallucination — unprompted MCQ generation
The model frequently converts simple completion prompts into multiple-choice quiz format with no instruction to do so. For example, asking "The plural of 'ox' is " caused the model to generate options A/B/C/D and a <think> reasoning block.

2. Factual knowledge gaps
The model failed to name Australia's capital (Canberra), instead generating a nonsensical pattern about distances between capitals. It also drifted completely off-topic when asked about UNESCO World Heritage Sites.

3. Taxonomic/scientific hallucination
The model reasoned that "birds are mammals that lay eggs" — a serious factual error showing confusion between biological categories.

---

What Fine-Tuning Data Would Fix These Errors?

1. Factual knowledge gaps

The model needs exposure to clean, factual question-answer pairs covering geography, science, history, and world knowledge.

Existing datasets to use:

• TriviaQA — 650K+ trivia QA pairs
• NaturalQuestions — real Google search questions with verified answers
• Wikidata triples — structured factual knowledge (country→capital, element→symbol, etc.)

Estimated size needed: ~50,000 examples to meaningfully reduce factual errors.

2. Format hallucination (unprompted MCQ generation)

The model needs examples of clean, direct completions — without MCQ scaffolding — so it learns that a plain sentence completion should be answered plainly.

How to assemble this:

• Take existing QA datasets and reformat them as completion-style prompts (e.g. "The capital of France is " → "Paris")
• Include negative examples during fine-tuning: show the model that generating A. B. C. D. options when none were requested is undesirable
• Filter out any training examples that contain unprompted MCQ structure

Estimated size needed: ~20,000 carefully curated completion pairs with no MCQ format in the target output.

3. Irregular grammar and linguistic exceptions

The model said "oxes" instead of "oxen", suggesting weak coverage of irregular English forms.

How to assemble this:

• Extract irregular plurals, verb conjugations, and exceptions from Wiktionary (freely available and structured)
• Augment with grammar textbook exercises and English language learning datasets
• CLMET3 corpus and similar linguistic resources

Estimated size needed: ~5,000–10,000 examples (this is a narrow domain so a smaller targeted dataset goes a long way).

---

Total Estimated Dataset Size

Error Type	Dataset Size Needed
Factual knowledge gaps	~50,000 examples
Format hallucination fix	~20,000 examples
Irregular grammar	~5,000–10,000 examples
Total	~75,000–80,000 examples

A dataset of this size should be sufficient for a LoRA fine-tune on the 0.8B model. Full fine-tuning would likely need 2–3x more data to avoid catastrophic forgetting.

---

Dataset Structure

Each row contains:

• input — the prompt given to the model
• expected_output — the correct answer
• model_output — what the model actually generated
• error_type — category of the failure