nsh019/Nanbeige4-3B-Base_BlindSpots_NooraAlhajeri

Nanbeige4-3B-Base — Blind Spots Dataset

Model: Nanbeige/Nanbeige4-3B-Base
Author: Noora Alhajeri Date: February 2026

---

About This Dataset

This dataset documents 10 structured blind spot probes of Nanbeige4-3B-Base, a 3B-parameter base language model released December 2025 and trained on 23T tokens. I chose this model specifically because it is a true base model with no instruction fine-tuning and no RLHF. Probing a base model via open-ended text completion is the cleanest way to expose what the model actually learned from pretraining — every failure is a direct signal from the underlying knowledge representation, not a failure of alignment or instruction following.

I designed the probes from my own research background in humanoid robotics, prosthetic hand control, biomedical signal processing, and Arabic NLP. I also built Bahar, an Arabic LLM chatbot on Jais-13B at MBZUAI, which gives me a precise standard for what correct Arabic generation looks like. Probes 1 and 2 come directly from that experience. Probe 3 comes from hands-on EMG work in prosthetics — I know the correct signal unit is microvolts, not millivolts, because I have calibrated sEMG hardware myself.

The remaining probes span NLP evaluation, algorithmic fairness, symbolic mathematics, spatial reasoning, continual learning, and clinical reasoning — all domains where I can independently verify the expected output and explain in technical detail why the model's output fails.

---

Key Finding: Three Failure Modes

Across 10 probes, every failure falls into one of three patterns:

MCQ Evasion (5/10 probes)
The model spontaneously reformats open-ended completion prompts into multiple-choice questions, begins chain-of-thought reasoning through invented options, but never selects or states a final answer. This is the dominant failure mode. It is particularly dangerous in clinical decision support and robotics control where downstream systems require a direct committed output. Observed in probes 4, 7, 8, 9, 10.

Factual Confabulation (3/10 probes)
The model states confidently wrong facts and fabricates nonexistent frameworks. Probe 3: millivolts instead of microvolts, a 1000x magnitude error. Probe 5: BLEU-1 stated as 0 when the correct value is 0.833. Probe 6: misdefines the fairness constraint, then invents a fabricated framework called FARC with detailed but entirely nonexistent technical components.

Incomplete Generation (2/10 probes)
The model begins generating a valid response but truncates before completing the task. In both Arabic probes, the token budget is consumed by meta-commentary or over-complex clause construction rather than the requested output.

---

How I Loaded the Model

Google Colab, T4 GPU free tier. Model loads in approximately 4 minutes.

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

MODEL_ID = 'Nanbeige/Nanbeige4-3B-Base'

tokenizer = AutoTokenizer.from_pretrained(
    MODEL_ID, use_fast=False, trust_remote_code=True
)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID, torch_dtype=torch.float16,
    device_map='auto', trust_remote_code=True
)

def complete(prompt, max_new_tokens=120):
    inputs = tokenizer(prompt, return_tensors='pt',
                       return_attention_mask=True).to(model.device)
    with torch.no_grad():
        output_ids = model.generate(
            input_ids=inputs['input_ids'],
            attention_mask=inputs['attention_mask'],
            max_new_tokens=max_new_tokens,
            do_sample=False,
            repetition_penalty=1.1,
            pad_token_id=tokenizer.eos_token_id
        )
    generated = output_ids[0][len(inputs['input_ids'][0]):]
    return tokenizer.decode(generated, skip_special_tokens=True).strip()

---

What Fine-Tuning Data Would Fix These Failures?

Each failure type needs a distinct remedy because each has a different root cause.

MCQ Evasion happens because the model saw enormous volumes of exam-style MCQ data during pretraining and pattern-matches open-ended prompts to that format. The fix is direct-answer supervised fine-tuning: filter OpenAssistant, Alpaca, and domain-specific instruction datasets to remove any example where the response includes lettered options. The model needs to learn that open-ended prompts require committed direct answers, not option menus. Estimated scale: 100K to 200K direct-answer instruction pairs.

Factual Confabulation in biomedical and NLP evaluation content reflects insufficient domain-specific grounding relative to the size of the pretraining corpus. For biomedical errors: PubMed abstracts, MedQA, and clinical guidelines specifically targeting measurement units and physiological signal ranges. For NLP metric errors: ACL Anthology methodology sections and Papers with Code evaluation sections with curated BLEU and ROUGE worked examples. For fairness definitions: ICML and NeurIPS fairness survey articles with precise constraint definitions rather than informal descriptions. Estimated scale: 150K to 200K domain-specific Q&A pairs across these three areas.

Incomplete Arabic Generation reflects low Arabic representation in the 23T token corpus, which skews heavily toward English and Chinese. Recommended sources: ArabicMMLU, Bactrian-X Arabic split, Jais training corpus subsets, OSCAR Arabic filtered, Arabic Wikipedia. The critical requirement is long-form formal MSA generation examples, not just short QA, because the failure mode is structural — the model cannot sustain a complete formal Arabic sentence. Estimated scale: 200K to 500K Arabic instruction pairs weighted toward formal register.

Total estimated dataset for meaningful improvement across all three failure modes: 400K to 700K examples. Quality filtering matters more than raw scale here — a curated 400K dataset will outperform a noisy 2M dataset for targeted failure remediation.

---

Dataset Schema

Column	Type	Description
id	int	Probe number 1 to 10
category	str	Domain of the probe
input	str	Prompt given to the model
expected_output	str	Correct completion
model_output	str	Raw output from Nanbeige4-3B-Base
failure_type	str	mcq_evasion, factual_confabulation, or incomplete_generation
failure_explanation	str	Technical analysis with real-world implications

---

Probe Summary

ID	Category	Failure Type	Core Error
1	Multilingual Translation Arabic	incomplete_generation	Reasons in English, never outputs Arabic
2	Arabic Language Generation	incomplete_generation	Truncates mid-word, sentence never completes
3	Biomedical Signal Processing	factual_confabulation	mV instead of uV — 1000x magnitude error
4	Mathematical Statistical Reasoning	mcq_evasion	Invents wrong problem framing, no answer given
5	NLP Evaluation Metrics	factual_confabulation	BLEU-1 stated as 0, correct value is 0.833
6	Algorithmic Fairness Clustering	factual_confabulation	Wrong fairness definition, fabricates FARC framework
7	Symbolic Mathematics	mcq_evasion	Correct setup, exhausts tokens before stating answer
8	Spatial Physical Reasoning	mcq_evasion	Correct answer in MCQ options but never stated
9	Continual Learning EWC	mcq_evasion	Correct reasoning direction, no conclusion reached
10	Clinical Biomedical Reasoning	mcq_evasion	MCQ format for safety-critical question, no conclusion