This is an efficiency-optimized version of the Deliberate-Qwen-3B model. While the standard version uses full-sentence reasoning, this Chain-of-Draft (CoD) version has been distilled to perform high-fidelity logic using minimalist "thought sketches."
🚀 Impact: The Efficiency Jump
Inference Speed: ~4.8x faster than standard CoT.
Token Efficiency: ~80% reduction in "Thinking" tokens.
Accuracy Retention: Maintained >95% performance on GSM8K-style logic benchmarks compared to the full-sentence model.
🧠 Why Chain-of-Draft?
In production environments, long internal monologues are expensive and slow. This model was trained using Rationale Compression Distillation to replace conversational filler with symbolic logic (e.g., 50*0.2=10 -> 50-10=40 instead of "First I will calculate 20% of 50...").
Example Output
Question: If I buy 3 shirts for $15 each and get a 10% discount, total?
Reasoning:<think> 3*15=45 -> 45*0.1=4.5 -> 45-4.5=40.5 </think> <answer> $40.50 </answer>
