Qwen2.5-R1-Minny-1.5B-v2

📐 Qwen2.5-R1-Minny-1.5B-v2

Qwen2.5-R1-Minny-1.5B-v2 is a highly experimental, mathematically optimized 1.5B parameter Small Language Model (SLM). It represents a completely custom, from-scratch implementation of a novel model merging paradigm: Sens-Stock Fusion.

By bridging the microscopic, gradient-based layer routing of Sens-Merging (Liu et al., 2025) with the macroscopic, flat-minimum geometric projection of Model Stock (Jang et al., 2024), this model achieves a Pareto-optimal fusion of pure mathematical reasoning, syntax generation, and conversational instruction-following.

🔬 The "Golden Triangle" Topology

This model was constructed without backpropagation by merging three highly specialized, homologous fine-tunes anchored to a foundational reasoning base:

• Anchor (Base): deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
• Expert A (Code): agentica-org/DeepCoder-1.5B-Preview
• Expert B (Math): RLinf/RLinf-math-1.5B
• Expert C (Chat/Structural): mobiuslabsgmbh/DeepSeek-R1-ReDistill-Qwen-1.5B-v1.1

🧮 The Sens-Stock Methodology

Standard model merging techniques (like SLERP or TIES) suffer from uniform coefficient assignments and arbitrary magnitude scaling. Sens-Stock solves both the directional and magnitude problems of task vector arithmetic mathematically:

1. Sens-Merging (The Direction): Using custom calibration subsets (MATH-500 Level 4/5, LeetCodeDataset, and Mixture-of-Thoughts), we extracted the first-order Taylor expansion approximations of parameter sensitivity via backpropagation. We combined this with $L_2$ cross-task logit alignment to generate temperature-scaled Softmax routing coefficients ($\sigma_i^l$) for every parameterized layer $l$. This guarantees that the Code model dominates syntax layers while the Math model dominates logical reasoning layers.
2. Model Stock (The Magnitude): Rather than uniformly scaling the routed task vector, we calculated the geometric angle ($\theta$) between the fine-tuned task vectors in high-dimensional space. We applied the Model Stock projection ratio $t = \frac{N \cos \theta}{1 + (N - 1)\cos \theta}$ (where $N=3$) to dynamically project the Sens-Merged vector to the boundary of the flat-minimum loss basin.

Final Fusion Equation: $$W_{final}^l = W_{base}^l + t^l \cdot \sum_{i=1}^{N} \sigma_i^l \cdot (W_{expert_i}^l - W_{base}^l)$$

💻 Usage & Prompt Format

Because this model is anchored to the DeepSeek-R1 distillation lineage, it utilizes the Qwen ChatML format and inherently relies on Chain-of-Thought (CoT) <think> tags for complex reasoning.

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "ForSureTesterSim/Qwen2.5-R1-Minny-1.5B-v2"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    torch_dtype=torch.bfloat16, 
    device_map="auto"
)

prompt = """Solve the following problem using Python.
First, explain the mathematical theory behind finding the nth Fibonacci number in O(log n) time using matrix exponentiation. 
Then, write a highly optimized Python function to implement it. Provide your reasoning inside a <think> block."""

messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs, 
        max_new_tokens=1024, 
        temperature=0.6,
        do_sample=True
    )

print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=False))