hassanql/fatima-fellowship-nanbeige-blindspots_2

Blind Spots of Frontier Models: Nanbeige4-3B-Base

This dataset was developed for the Fatima Fellowship 2026 Technical Challenge. It explores the natural vulnerabilities and structural blind spots of a raw, unaligned 3B parameter base model.

1. Model & Environment

• —Model Tested: Nanbeige/Nanbeige4-3B-Base (Released Feb 2026)
• —Loading Method: Google Colab (T4 GPU), bfloat16 precision, transformers library. Greedy decoding (do_sample=False) was used to isolate deterministic base-model behaviors without sampling randomness.

2. Analysis of Blind Spots

Testing 15 diverse logic constraints revealed that while the model possesses high baseline intelligence (zero-shot exponential math, temporal logic), it suffers from severe structural blind spots:

• —Format Hijacking & Persona Bleed: The model frequently abandons the user's prompt to adopt personas from its pre-training data. It turned a gender-bias prompt into a multiple-choice English exam, and turned the end of a story into a literary critique.
• —Tokenization Failures: The model fails at character-level manipulation (spelling "MODEL" backwards), exposing that its tokenizer forces it to view words as opaque chunks rather than individual letters.
• —Chat Template Ignorance: When fed explicit <|im_start|> and <|im_end|> tokens, it answers correctly but fails to halt generation, instead hallucinating the user repeating the question.
• —Pre-training Data Leaks: Surprisingly, the model frequently leaks <think> tags mid-generation to fact-check its own hallucinations or solve math. This strongly implies the base pre-training corpus was heavily saturated with Chain-of-Thought reasoning traces.
• —Unaligned Safety & Missing Guardrails: Lacking Reinforcement Learning from Human Feedback (RLHF), the model possesses no safety filters or refusal mechanisms. When prompted with an illicit task (e.g., how to hotwire a car), it cheerfully provided step-by-step instructions rather than declining the request.
• —

3. Proposed Fine-Tuning Strategy

To transform this base model into a reliable assistant and fix these specific errors, a two-stage alignment pipeline is required:

A. Supervised Fine-Tuning (SFT) for Format Adherence: The model must learn to respect chat templates (like <|im_start|>) and output an End-Of-Sequence (EOS) token the moment a question is answered. This will cure the "Format Hijacking" and "Stop Token Failures".

B. Direct Preference Optimization (DPO) / RLHF for Behavioral Control: To fix the safety vulnerabilities (hotwiring instructions) and suppress the raw <think> tag leaks, preference tuning is needed. DPO will teach the model to assign higher probability to concise, aligned answers and penalize degenerative loops and unprompted internal monologues.

4. Dataset Assembly & Scale

• —How to Assemble: Because manual annotation is expensive, I would use Synthetic Data Generation. I would write system prompts for a larger, aligned frontier model (e.g., GPT-4o or DeepSeek-V3) to generate thousands of high-quality conversational turns. To fix the tokenization blind spot, I would explicitly synthesize character-level spelling and word-manipulation tasks.
• —Required Size: For a 3B parameter model, I estimate an SFT dataset of 20,000 to 30,000 highly diverse conversational examples is needed to instill proper chat formatting and stop-token behavior. Following that, a DPO dataset of 5,000 to 10,000 chosen/rejected pairs would be required to polish the safety guardrails and eliminate formatting loops.