hari31416/grug-reasoning-data-and-benchmarks

Grug Reasoning Datasets and Benchmark Results

This repository contains all training datasets, preference pairs, raw model generation logs, and empirical benchmark results for the Grug reasoning research project (spanning both 1.5B and 7B models).

Repository Contents

├── data/
│   ├── 1.5b/
│   │   ├── it-1/                 # Iteration 1 SFT data and compressed traces
│   │   └── it-2/                 # Iteration 2 scaled SFT data
│   ├── 7b/
│   │   ├── sft/                  # 7B SFT training and validation splits
│   │   └── dpo/                  # 7B DPO preference pairs (chosen vs rejected)
│   ├── sft/                      # Root pointer to 7B SFT
│   └── dpo/                      # Root pointer to 7B DPO
├── benchmarks/
│   ├── deepseek-r1-1.5b/
│   │   ├── it-1/                 # 1.5B Iteration 1 GSM8K evaluation JSONs
│   │   └── it-2/                 # 1.5B Iteration 2 GSM8K evaluation JSONs
│   └── deepseek-r1-7b/
│       ├── baseline/gsm8k.json   # Full 1,319 GSM8K evaluations (Base 7B)
│       ├── finetuned/gsm8k.json  # Full 1,319 GSM8K evaluations (SFT 7B)
│       └── dpo/gsm8k.json        # Full 1,319 GSM8K evaluations (DPO 7B)
└── reports/
    ├── deepseek-r1-1.5b/
    │   ├── it-1/                 # 1.5B Iteration 1 dashboard plots and report PDF
    │   └── it-2/                 # 1.5B Iteration 2 dashboard plots and markdown report
    └── deepseek-r1-7b/
        ├── benchmark_comparison_dashboard.png
        ├── loss_plot.png
        └── BENCHMARK_REPORT.md

Benchmark Results

DeepSeek-R1-7B (Full GSM8K Test Split, 1,319 Samples)

Model Variant	Samples	Accuracy	Format Compliance	Mean Thinking Tokens	Mean Answer Tokens	Mean Latency
Base Model (7B)	1,319	75.97%	99.85%	122.5	160.4	6.09s
SFT Adapter (7B)	1,319	72.18%	94.62%	107.7	107.3	6.75s
DPO Adapter (7B)	1,319	75.44%	99.85%	122.3	162.1	6.39s

DeepSeek-R1-1.5B (Apple Silicon Experiments)

Configuration	Accuracy	Mean Thinking Tokens	Mean Total Tokens	Mean Latency	Format Compliance
Base Normal (1.5B)	64.9%	219.0	477.4	0.88s	96.6%
FT Normal (1.5B)	66.0%	156.2	389.3	0.73s	98.9%
FT Regularized (1.5B)	54.6%	135.0	214.7	0.61s	98.2%

Visualizations

7B Benchmark Comparison Dashboard

7B DPO Training Loss Curve

1.5B Iteration 2 Dashboard

How to Load via Hugging Face Datasets

from datasets import load_dataset

# Load 7B SFT dataset
sft_7b = load_dataset("hari31416/grug-reasoning-data-and-benchmarks", data_files="data/7b/sft/train.jsonl", split="train")

# Load 7B DPO preference pairs
dpo_7b = load_dataset("hari31416/grug-reasoning-data-and-benchmarks", data_files="data/7b/dpo/train.jsonl", split="train")

# Load 1.5B SFT dataset
sft_15b = load_dataset("hari31416/grug-reasoning-data-and-benchmarks", data_files="data/1.5b/it-2/train.jsonl", split="train")

Related Resources

• Model LoRA Adapters: hari31416/deepseek-r1-grug-adapters
• GitHub Project Repository: Hari31416/qwen-grug-finetune