Oluwashayo/qwen3.5-0.8b-base-blindspots

Qwen3.5-0.8B-Base Blindspots

This dataset documents the observed reasoning and behavioral blindspots of the Qwen3.5-0.8B-Base model during evaluation experiments.

---

1. Model Tested

Model Name: Qwen/Qwen3.5-0.8B-Base
Architecture: Unified Vision-Language (Gated DeltaNet hybrid)
Parameters: 0.8 Billion
Release Date: March 2026

---

2. Loading Procedure

The model was evaluated using a Google Colab environment with a T4 GPU.

To ensure efficient memory usage, the model was loaded using bfloat16 precision.

import transformers
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

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

try:
    print("Loading tokenizer...")
    tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)

    print("Loading model...")
    model = AutoModelForCausalLM.from_pretrained(
        model_id,
        torch_dtype=torch.bfloat16,
        device_map="auto",
        trust_remote_code=True
    )
    print("Model loaded successfully!")
except Exception as e:
    print(f"\nERROR: {e}")

---

3. Experimentation & Error Analysis

As a base model, Qwen3.5-0.8B-Base demonstrates high raw reasoning capabilities (visible via its internal tags) but struggles with instruction alignment and output grounding.

Key Blind Spots Identified:

• Instruction Following: The model often fails to respect constraints such as "no explanation" or "exactly 2 sentences."

• Logical "Trap" Riddles: It frequently falls for pattern-based traps (e.g., naming the third child "June" instead of "John" due to month-name associations).

• Hallucination: It generates specific, plausible-sounding details (like scores for a 2025 World Cup) for events that have not occurred or are non-existent.

• Tokenization Issues: Standard counting tasks (e.g., counting 'r' in "cherry") trigger repetitive looping and incorrect counts.

---

4. Fine-Tuning Strategy

To resolve these errors, the model requires Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF).

• Dataset Recommendation: I recommend utilizing the Open-Orca or ShareGPT datasets to improve conversational flow and instruction following.

• Dataset Assembly: For logic-specific errors, I would assemble a synthetic dataset of 15,000–20,000 "CoT-to-Answer" pairs where a larger teacher model (like Qwen2.5-72B) corrects the logical leaps observed in this 0.8B version.

• Focus Area: Specific focus should be placed on Negative Constraint training (e.g., "Do not use X word") to improve the model's attention to user-defined boundaries.