Trouter-Terminus-20b

---

license: apache-2.0 datasets:

• HuggingFaceFW/finewiki metrics:
• accuracy
• character pipeline_tag: text-generation library_name: allennlp tags:
• agent
• code

---

Trouter-Terminus-20b

<div style="display:flex;justify-content:center;align-items:center;height:100%;"> <svg viewBox="0 0 200 200" width="320" height="200" xmlns="http://www.w3.org/2000/svg"> <path d=" M 18 120 C 18 96, 30 96, 30 120 C 30 144, 54 144, 54 120 C 54 72, 78 72, 78 120 C 78 168, 102 168, 102 120 C 102 72, 126 72, 126 120 C 126 168, 150 168, 150 120 C 150 96, 162 96, 182 120 " fill="none" stroke="White" stroke-width="16" stroke-linecap="round" stroke-linejoin="round" /> </svg> </div> <div align="center">

</div>

Introduction

Trouter-Terminus-20b is the second generation in the Trouter model series, building upon the foundation of Trouter-20b with significant enhancements. This update maintains the model's original capabilities while addressing issues reported by users, including:

• Enhanced Reasoning: Improved performance in logical reasoning and complex problem-solving tasks
• Superior Code Generation: Advanced capabilities in code understanding, generation, and debugging across multiple programming languages
• Optimized Performance: Better efficiency and accuracy across diverse benchmarks

Benchmark Comparisons

Performance Improvements

Key Improvements

• Reasoning: Significant gains in mathematical reasoning (GSM8K: +7.4) and advanced reasoning tasks (GPQA-Diamond: +6.5)
• Code Generation: Major improvements in code completion (HumanEval: +10.4) and competitive programming (Codeforces: +164 rating points)
• General Knowledge: Enhanced performance on multi-task language understanding (MMLU-Pro: +5.8)

Model Details

• Parameters: 20 billion
• Architecture: Transformer-based
• Training: Continued training on Trouter-20b with enhanced datasets
• Context Length: Extended context window for better long-form understanding
• License: MIT

Use Cases

• Advanced code generation and debugging
• Complex reasoning and problem-solving
• Multi-step task execution
• Technical documentation generation
• Data analysis and interpretation

Getting Started

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("OpenTrouter/Trouter-Terminus-20b")
model = AutoModelForCausalLM.from_pretrained("OpenTrouter/Trouter-Terminus-20b")

prompt = "Write a Python function to calculate fibonacci numbers:"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=200)
print(tokenizer.decode(outputs[0]))

Limitations

• May occasionally generate incorrect information
• Performance varies based on prompt quality
• Not intended for production use without proper evaluation
• May reflect biases present in training data

Citation

@misc{trouter-terminus-20b,
  title={Trouter-Terminus-20b: Enhanced Reasoning and Code Generation},
  author={Andrew Kellen},
  year={2024},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/OpenTrouter/Trouter-Terminus-20b}}
}

Acknowledgments

Built upon Trouter-Terminus-20b with community feedback and extensive testing. Special thanks to all users who provided valuable insights for this iteration.