aries-reasoning-1.5b

Aries Reasoning 1.5B

Aries Reasoning 1.5B is a custom 1.54B parameter model built on the Aries Base Architecture, designed for structured, stable chain-of-thought reasoning and low-VRAM inference.

🔥 Reasoning Token Format

Aries uses 4 explicit reasoning tokens, escaped so HuggingFace renders them correctly:

• \<think>
• \<context>
• \<answer>
• \<end>

These tokens allow the model to separate:

• internal reasoning steps
• external user-facing answers
• injected context
• termination boundaries

🚀 Features

• Explicit reasoning tokens for controllable CoT
• CPU-offloaded optimizer during training
• Gradient checkpointing enabled
• Only 10–11GB GPU RAM needed for inference
• Trained on GSM8K-style structured reasoning data
• Strong jailbreak resistance (does not leak CoT when prompted)

🧪 Usage Example

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("ziadrone/aries-reasoning-1.5b")
model = AutoModelForCausalLM.from_pretrained(
    "ziadrone/aries-reasoning-1.5b",
    torch_dtype="bfloat16",
    device_map="auto"
)

prompt = "<think> What is 123 + 456?"
inputs = tok(prompt, return_tensors="pt").to("cuda")
out = model.generate(**inputs, max_new_tokens=80)

print(tok.decode(out[0]))