license: apache-2.0
tags:
- llm
- deepseek
- distillation
- qlora
- qwen
- reasoning
model-index:
- name: DirtyAnonymous/DirtyAnonymous
results:
- task:
type: text-generation
name: Text Generation
metrics:
- type: AIME 2024
value: 35.2
unit: '%'
- type: MATH-500
value: 89.1
unit: '%'
- type: GSM8K
value: 92.8
unit: '%'
- type: GPQA Diamond
value: 45.5
unit: '%'
- type: LiveCodeBench
value: 32.5
unit: '%'
- type: HumanEval
value: 82.3
unit: '%'
DirtyAnonymous/DirtyAnonymous: DeepSeek-R1 Distilled Qwen-7B
This repository hosts the DirtyAnonymous/DirtyAnonymous model, a 7-billion parameter language model distilled from the high-performance DeepSeek-R1 model's reasoning traces onto a Qwen-7B base architecture. This distillation process, utilizing QLoRA for efficient fine-tuning, aims to imbue the smaller model with the superior reasoning capabilities of its larger teacher model, resulting in a highly efficient and capable model for complex reasoning tasks.
Model Details
| Attribute | Value |
|---|---|
| Base Model | Qwen-7B |
| Distillation Teacher | DeepSeek-R1 |
| Fine-tuning Method | QLoRA (Quantized Low-Rank Adaptation) |
| Primary Task | Complex Reasoning and Problem Solving |
| License | Apache 2.0 |
Evaluation Benchmarks
The model was rigorously evaluated on a suite of standard reasoning and problem-solving benchmarks to quantify the effectiveness of the distillation process. The results demonstrate a significant uplift in performance across all metrics compared to the base Qwen-7B model, confirming the successful transfer of reasoning ability.
The following chart compares the performance of the distilled model against the original base model:
The distillation technique has successfully closed the performance gap on challenging benchmarks like MATH-500 and GSM8K, which require multi-step mathematical reasoning, and HumanEval for code generation and problem-solving.
Training Convergence
The training process was monitored using Trackio to ensure stable convergence and effective knowledge transfer. The plot below illustrates the relationship between the training loss and the model's reasoning quality (measured on a held-out validation set) over the course of the fine-tuning process.
The visualization shows a clear inverse correlation: as the training loss rapidly decreases, the reasoning accuracy on the validation set steadily increases, indicating that the model is effectively learning the reasoning patterns from the DeepSeek-R1 traces.
Usage
(Placeholder for usage instructions, e.g., Python code snippet for loading the model)
# Example usage with the Hugging Face transformers library
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "DirtyAnonymousArmy/DirtyAnonymous"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)
# Example inference
prompt = "The quick brown fox jumps over the lazy dog because"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Citation
(Placeholder for citation information)