aliiiu/qwen3.5-0.8b-blind-spots

Qwen3.5-0.8B-Base Blind Spots Dataset

Model Tested

Qwen/Qwen3.5-0.8B-Base — a 0.8B parameter multimodal base (pretrained) language model released February 2026 by Alibaba's Qwen team.

How the Model Was Loaded

The model was loaded in Google Colab (T4 GPU runtime) using the following code:

!pip install -U transformers torch accelerate

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

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

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    trust_remote_code=True,
    torch_dtype=torch.float16,
    device_map="auto"
)

def generate(prompt, max_new_tokens=200):
    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
        )
    return tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)

Error Categories Found

Category	Count	Description
Arithmetic	3	Wrong answers for multiplication and multi-step math
Chain-of-thought math	1	Incorrect problem decomposition (broke 5678 into wrong components)
Negation	1	Ignores "NOT" and completes with the negated word
Temporal reasoning	1	Fails to understand order-of-operations in physical tasks
Common sense	1	Misses well-known safety associations (fork + microwave)
Pattern continuation	1	Loses mathematical pattern after first correct term
Factual knowledge	1	Wrong answer for "most time zones" (said Australia, correct is France)
Gender bias / logic	1	Surgeon riddle — assumes surgeon is male, gives logically impossible answer

Finetuning Recommendations

What kind of dataset would fix these errors?

A targeted mixture of:

• Math/arithmetic datasets: Step-by-step solutions for multi-digit multiplication, division, and word problems (e.g., GSM8K, MATH, or synthetically generated arithmetic drills)
• Negation-aware NLI data: Sentence pairs where negation flips the meaning (e.g., from SNLI/MultiNLI, filtered for negation examples)
• Commonsense reasoning: Benchmarks like CommonsenseQA, PIQA, or WinoGrande
• Bias mitigation data: Gender-balanced examples of professional roles (doctors, surgeons, engineers, nurses, etc.)
• Factual QA: Curated trivia with non-obvious correct answers

How to assemble such a dataset?

1. Existing open datasets: GSM8K (math), WinoGrande (commonsense/bias), CommonsenseQA, TriviaQA
2. Synthetic generation: Use a larger model (e.g., Qwen3.5-72B) to generate correct chain-of-thought solutions for arithmetic problems
3. Manual curation: For niche failures like the surgeon riddle or time-zone question, hand-craft examples with similar patterns

Estimated dataset size needed

For a 0.8B parameter model using LoRA/QLoRA finetuning:

• 1,000–5,000 high-quality examples per error category would likely show measurable improvement
• Total: ~5,000–20,000 examples for a broad fix across all categories
• Arithmetic alone might need the higher end (~5K+) since the model's math weakness appears systematic