afalaudn/nanbeige-3b-blind-spots

Blind Spots Evaluation: Nanbeige/Nanbeige4-3B-Base

Model Tested

• Model name: Nanbeige4-3B-Base
• Parameter count: 3B
• Architecture: LlamaForCausalLM
• Release date: 06 December 2025
• Confirmation: This is a pure base model with no chat template applied. It requires manual completion or few-shot prompting for structured tasks.

How to Load the Model

Include this exact working code:

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_name = "Nanbeige/Nanbeige4-3B-Base"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
tokenizer.pad_token = tokenizer.eos_token
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True,
)

def generate(prompt, max_new_tokens=2048):
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=max_new_tokens,
            do_sample=False,
            pad_token_id=tokenizer.eos_token_id,
            repetition_penalty=1.1,
        )
    new_tokens = outputs[0][inputs["input_ids"].shape[1]:]
    return tokenizer.decode(new_tokens, skip_special_tokens=True)

Evaluation Platform

• Environment: Modal.com with NVIDIA L4 GPU (24GB VRAM)
• Settings: Greedy decoding, max_new_tokens=2048, repetition_penalty=1.1
• Scope: 200 prompts across 10 categories (partial evaluation of 68 prompts in this version)

Interesting Finding

Unexpected <think> tags appeared in the base model's output even though it had no explicit reasoning or RLHF training in the public description. This suggests that the pre-training data might have included a significant amount of chain-of-thought data or web crawls of model outputs (like DeepSeek's outputs) which the model learned to mimic.

Dataset Structure

Column	Description
id	Unique identifier for the prompt
category	Type of test (negation, temporal, logic, etc.)
input_prompt	The exact prompt sent to the model
expected_output	The objectively correct answer
model_output_final	The final answer extracted from the model
model_output_thinking	The chain-of-thought or thinking process generated
error_type	Classification of the error (factual, temporal, etc.)
notes	Explanation of why the model failed

Blind Spots Found

category	errors found	total tested	error rate	description of pattern
negation	10	10	100%	Failed to ignore negative constraints; entered repetition loops.
temporal	15	20	75%	Confused date relative offsets (3 days ago from Wednesday).
common_sense	12	20	60%	Hallucinated additional context instead of simple inference.
formatting	18	18	100%	Completely failed structured ordering; entered endless loops.

Why Does the Model Fail? (Root Cause Analysis)

• Tokenization & Context: The model likely struggles with specific relative markers in temporal logic due to how it tokens sequence dependencies.
• Pre-training Distribution: A strong bias towards Chinese-centric data might make performance on English-specific nuances (like "NOT" items) less robust.
• Lack of Chat-Tuning: As a base model, it defaults to completion. Without a chat template, it "completes" the task by hallucinating a whole dialogue or additional questions.

Fine-tuning Recommendations

Recommended Datasets to Fix These Errors

• arithmetic/math: GSM8K, MATH dataset
• logical reasoning: LogiQA, ReClor, ProofWriter
• Indonesian language: Indonesian SQuAD, IndoNLU
• factual: FEVER, TriviaQA

How to Assemble Such a Dataset

1. Existing Benchmarks: Subsample high-quality reasoning logs from existing datasets.
2. Synthetic Generation: Use LLM to generate complex "negation" prompts and verify with a separate "critic" model.
3. Human Annotation: Focus on edge cases where models typically hallucinate, specifically in temporal multi-step reasoning.

Estimated Dataset Size Needed

According to the LIMA paper, 1000 carefully curated, high-quality examples can be competitive with 50K noisy examples. For this 3B model, a targeted SFT dataset of 2000-5000 examples focusing on the specific blind spots (negation, formatting) using LoRA or full fine-tuning would likely yield significant improvements.