phi4-mini-ultra-reasoning-GGUF

model_name: "Phi-4-Mini-Ultra-Reasoning (The Lobotomized Edition)"

critical_warning: > DO NOT DOWNLOAD. This model is fundamentally broken. It suffers from catastrophic forgetting, EOS (End of Sentence) token blindness, and severe attention loops. Do not use this for production, development, or basic arithmetic. It is uploaded purely for archival, educational, and morbidly curious testing purposes.

intended_purpose: base_model: "microsoft/Phi-4-mini-reasoning" goal: > To create a highly capable, localized reasoning agent fine-tuned on some of the hardest reasoning, coding, and logical datasets available (Opus 4.6, Gemini 3.1 Pro Hard, DeepResearch, Nemotron Terminal Corpus, etc.).

actual_state_autopsy: description: "Instead of a hyper-smart agent, this fine-tuning run accidentally gave the model permanent brain damage." symptoms: - "Ignores the user's prompt entirely." - "Enters an infinite <|thought|> loop (e.g., repeating 'The user is simply greeting me' 500 times)." - "Spontaneously generates random XML, Java, or C++ code regardless of context." - "Speaks in a 'multilingual word salad' (mixing English, Russian Cyrillic, and code symbols)."

fatal_flaws:

• parameter: "Learning Rate (LR)" value: "2e-4" impact: > Acts as a sledgehammer on a delicate architecture. Caused Catastrophic Forgetting, violently overwriting the model's base knowledge of English grammar, formatting, and logic.
• parameter: "Sequence Length (SEQ_LEN)" value: "512" impact: > Suffocated the context window. Cut off massive reasoning chains mid-sentence. The model never saw an EOS (End of Sentence) token or a <|end|> tag during training, causing it to completely un-learn how to stop generating text.

gguf_frankenstein_saga: description: "The GGUF files were forged through a dark series of hacks because llama.cpp refused to compile the broken weights." sequence_of_events: 1: "The Phantom Tokenizer: The Hugging Face repo contained a corrupted, 0-byte tokenizer.model file causing compiler crashes." 2: "The Bait and Switch: Deleted the corrupted Phi-4 tokenizer and secretly swapped in a tokenizer.model stolen from the older microsoft/phi-2 repository." 3: "The Result: The compiler accepted the dummy file and quantized the weights, resulting in structural GGUFs with a broken intellect."

available_files: description: "For AI researchers studying model collapse, over-fitting, or attention mechanism failures." files: - name: "phi4-mini-f16.gguf" notes: "The massive, uncompressed broken brain." - name: "phi4-mini-q8_0.gguf" notes: "8-bit integer." - name: "phi4-mini-q4_K_M.gguf" notes: "The standard, recommended format for observing the hallucinations." - name: "phi4-mini-q4_K_S.gguf" notes: "Slightly smaller 4-bit." - name: "phi4-mini-q2_K_M.gguf" notes: "2-bit—somehow even worse than the others."

lessons_learned:

• "Drop your Learning Rate: Use 1e-5 or 5e-6 for Phi-4 family."
• "Expand your Sequence Length: Ensure SEQ_LEN is 2048 or 4096 to capture full reasoning blocks."
• "Verify your EOS Tokens: Ensure your dataset mapping function appends <|end|> to every single prompt."