⚛️ Atomight-2-1.5B-Thinking

A Deep-Reasoning Small Language Model Optimized for Sequential Logic Chains

📌 Model Overview

Atomight-2-1.5B-Thinking is a specialized, compact reasoning model built on top of a 1.5B parameter core architecture. Engineered explicitly for users operating on constrained hardware environments (such as a free Google Colab T4 instance), Atomight-2 utilizes an explicit internal <think>...</think> scratchpad layout. It dynamically breaks down complex mathematical, logical, and structural prompts before committing to a final conclusion.

🚀 Key Highlights

• Hardware Democratic: High-tier deep reasoning accessible on consumer-grade hardware and free cloud compute tiers.
• Structured Scratchpad: Generates native, visible reasoning pathways natively formatted for transparent auditing.
• Chat-Template Native: Tailored directly for ChatML system configurations.

---

📊 Evaluation & Benchmark Results

Atomight-2 was subjected to a high-volume statistical evaluation matrix across core logic paradigms, matching up against premier industry baselines in the 1B–4B small language model class.

Official Performance Breakdown

The model displays exceptional specialization spikes in structured mathematical deduction, rivaling or outperforming significantly larger parameters classes on core numerical strings.

Benchmark	Paradigm	Atomight-2-1.5B-Thinking	Qwen-2-1.5B-Instruct	Phi-3-mini (3.8B)	Llama-3.2-3B-Instruct
GSM8k	Math Logical Chains	80.1%	71.0%	82.5%	73.1%
ARC-C	Core Reasoning	88.5%	82.3%	84.9%	83.3%
MMLU	General Knowledge	63.2%	56.7%	68.8%	61.1%

⚠️ Evaluation Insight: While Atomight-2 exhibits class-leading spikes on core textual logic and mathematical proofs, it experiences a classic reasoning tradeoff. On abstract matrix-grid visual transformation evaluations (like ARC-AGI 2), it drops to a baseline floor of 0.00%. This cognitive bottleneck highlights an instruction deficit in translating spatial imagery into basic structural text tokens—a major priority slated for the next architecture generation.

---

💻 Quickstart & Inference Code

To deploy Atomight-2 cleanly without encountering text-truncation errors inside the internal reasoning blocks, execute the generation using the official structured chat template format.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

MODEL_ID = "NovatasticRoScript/Atomight-2-1.5B-Thinking"

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

# Structure conversational dialog into ChatML framework
messages = [
    {"role": "user", "content": "A retailer buys shirts for $15 and sells them for $25. What is the total profit on 12 shirts?"}
]

templated_input = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(templated_input, return_tensors="pt").to("cuda")

print("🧠 Generating Reasoning Sequence:")
outputs = model.generate(
    **inputs, 
    max_new_tokens=768, # Plentiful headroom required for deep-thinking scratchpads
    temperature=0.1,
    do_sample=False,
    pad_token_id=tokenizer.eos_token_id
)

print(tokenizer.decode(outputs[0], skip_special_tokens=False))