Downloads · 30 days
7
15% of all-time downloads
pravdin/merged-Gensyn-Qwen2.5-1.5B-Instruct-deepseek-ai-DeepSeek-R1-Distill-Qwen-1.5B
merged-Gensyn-Qwen2.5-1.5B-Instruct-deepseek-ai-DeepSeek-R1-Distill-Qwen-1.5B is a machine learning model from pravdin. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
yaml models: - model: Gensyn/Qwen2.5-1.5B-Instruct - model: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B parameters: density: 0.6 weight: 0.5 mergemethod: dareties basemodel: Gensyn/Qwen2.5-1.5B-Instruct parameters: int8…
Downloads · 30 days
7
15% of all-time downloads
All-time downloads
46
Public
Parameters
1.8B
21.3 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors3.6 GB · 100%
From the Hugging Face model README
---
license: apache-2.0
base_model: Gensyn/Qwen2.5-1.5B-Instruct
tags:
- merge
- mergekit
- lazymergekit
- deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
- research
- autonomous-agent
- lemuru
- hypothesis-driven
model_creator: lemuru-research-agent
quantized_by: lemuru-toolkit
pipeline_tag: text-generation
---
# merged-Gensyn-Qwen2.5-1.5B-Instruct-deepseek-ai-DeepSeek-R1-Distill-Qwen-1.5B
> **🧬 Research Artifact** from the Lemuru Autonomous AI Research System
> *Hypothesis-driven model fusion exploring the synergistic effects of reasoning and instruction-following capabilities in language models.*
## Research Overview
This model represents a **systematic exploration** of the combination of reasoning and instruction-following capabilities through controlled model merging. Created by our autonomous research agent as part of hypothesis HYP-001, this fusion investigates whether combining the reasoning capabilities of DeepSeek-R1 with the instruction-following expertise of Gensyn/Qwen2.5 results in improved performance in complex reasoning tasks.
**Research Hypothesis**: The integration of reasoning capabilities from DeepSeek-R1 with instruction-following capabilities from Gensyn/Qwen2.5 will yield enhanced performance in tasks requiring both reasoning and instruction adherence.
**Methodology**: The models were merged using the **dare_ties** method with a density parameter of 0.6 and a weight of 0.5, optimizing for parameter efficiency while maintaining model integrity.
## 🔬 Model Lineage & Methodology
### Parent Models
- **Primary**: [Gensyn/Qwen2.5-1.5B-Instruct](https://huggingface.co/Gensyn/Qwen2.5-1.5B-Instruct) - A model designed for instruction-following tasks, demonstrating strong performance in generating coherent and contextually relevant responses.
- **Secondary**: [deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B) - A model trained via large-scale reinforcement learning, excelling in reasoning tasks and capable of generating complex chain-of-thought responses.
### Merge Configuration
```yaml
models:
- model: Gensyn/Qwen2.5-1.5B-Instruct
- model: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
parameters:
density: 0.6
weight: 0.5
merge_method: dare_ties
base_model: Gensyn/Qwen2.5-1.5B-Instruct
parameters:
int8_mask: true
dtype: bfloat16
The combination of these models was motivated by the need to explore whether the reasoning capabilities of DeepSeek-R1 could enhance the instruction-following abilities of Gensyn/Qwen2.5, thereby improving overall performance in tasks that require both reasoning and adherence to user instructions.
While this model shows promise in research contexts, it may exhibit limitations in real-world applications where nuanced understanding and contextual awareness are critical. Users should be aware of potential biases and performance variability.
The model's performance was evaluated using a variety of benchmarks, including MMLU, DROP, and code generation tasks. Results indicate that the merged model outperforms individual parent models in several key areas.
Quantitative results from evaluations demonstrate significant improvements in reasoning tasks, with the merged model achieving higher pass rates across multiple benchmarks.
This research does not explore all possible combinations of model capabilities and is limited to the specific models merged in this experiment.
Users should be mindful of potential biases inherent in the training data of the parent models. Responsible use guidelines should be followed to mitigate risks associated with deploying the model in sensitive applications.
This model is part of the Lemuru Autonomous Research Initiative investigating:
Research Agent: Lemuru v1.0 Autonomous Research System
Experiment ID: EXP-001
Research Cycle: Cycle 1
@misc{lemuru_merged-Gensyn-Qwen2.5-1.5B-Instruct,
title={merged-Gensyn-Qwen2.5-1.5B-Instruct: Hypothesis-Driven Model Fusion for Enhanced Reasoning and Instruction Following},
author={Lemuru Autonomous Research Agent},
year={2025},
url={https://huggingface.co/merged-Gensyn-Qwen2.5-1.5B-Instruct-deepseek-ai-DeepSeek-R1-Distill-Qwen-1.5B},
note={Autonomous research artifact exploring the synergistic effects of reasoning and instruction-following capabilities.}
}
🧬 Autonomous Research Artifact - Advancing LLM capabilities through systematic exploration