🦊 Fox1.4 - Reasoning Specialist

Fox1.4 is Fox1.3's successor, trained on combined data from math, logic, knowledge, and code reasoning tasks.

Performance

Custom Benchmark (10 questions):

• ✅ All tasks: 100%
• Penguin exception logic: ✅
• $1.10 riddle: ✅
• Math (2+2, 15+27, 100/4, 7*8): ✅
• Knowledge (France, Jupiter): ✅
• Code (is_even): ✅

Estimated MMLU Score: ~40-50%

Architecture

• Base Model: Qwen2.5-0.5B (merged with LoRA adapter)
• Training: Combined data from 4 expert domains
• Parameters: ~900M
• Format: Full merged model (safetensors)

Usage

Ollama

ollama pull teolm30/fox1.4
ollama run fox1.4

Python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("teolm30/fox1.4")
tokenizer = AutoTokenizer.from_pretrained("teolm30/fox1.4")

inputs = tokenizer("Your question", return_tensors="pt")
outputs = model.generate(**inputs)
print(tokenizer.decode(outputs[0]))

HuggingFace Inference

Click the "Use this model" button above to run inference directly on HuggingFace.

Comparison

Feature	Fox1.3	Fox1.4
Base	Qwen2.5-0.5B	Qwen2.5-0.5B
Training	LoRA	Merged LoRA
Format	GGUF	Safetensors
Custom Benchmark	100%	100%
Size	~1 GB	~1 GB

Model Details

• Parameters: ~900M
• Context Length: 16K
• Quantization: None (full bf16)
• Hardware: Runs on CPU or GPU

---

Fox1.4 — focused reasoning at its best.