Ndong21/gwen2.5-3B-blind-spots

Blind Spots Evaluation

Model Tested

Qwen/Qwen2.5-3B
https://huggingface.co/Qwen/Qwen2.5-3B

---

Code to Load the Model

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "Qwen/Qwen2.5-3B"

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

def generate(prompt, max_tokens=100):
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=max_tokens,
            temperature=0.7,
            do_sample=True
        )
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

Proposed dataset type

The dataset should prioritize Chain-of-Thought (CoT) traces that demonstrate the "how" before the "what." By including counter-intuitive scenarios, permutation training for famous riddles, and formal logic proofs, the model learns to prioritize underlying rules over statistical memorization.

How to find data

Since high-quality logic is scarce online, use a synthetic pipeline involving code-verified math problems and "Teacher-Student" inversion. This involves using a more capable model to generate reasoning paths for simple logic tasks, then auditing them via human experts to ensure the ground truth is flawless.

Size of dataset

A curated set of 50,000 to 100,000 examples is sufficient to "unlock" reasoning behaviors, provided each logical blind spot is represented by at least 1,000 distinct variations to prevent the model from simply memorizing the answers.