CoolFace
Datasetpublic

GlimmaryKarl/DualBlind

GlimmaryKarl/DualBlind Curated Frontier Reasoning and Direct Preference Optimization (DPO) Dataset Generated from Double-Blind Multi-Agent Arena Evaluations. This dataset was generated using the DualBlind AI Benchmark Arena. In this setup, two independent frontier AI models engage in multi-turn double-blind dialogue to solve extreme-difficulty benchmark problems, verifying their peer's proofs, raising counter-examples, and reaching mathematical consensus. Dataset… See the full description on the dataset page: https://huggingface.co/datasets/GlimmaryKarl/DualBlind.

sourceHugging Faceapache-2.0updated 15d agoView on Hugging Face
0likes196downloads
Dataset Card

GlimmaryKarl/DualBlind

Curated Frontier Reasoning and Direct Preference Optimization (DPO) Dataset Generated from Double-Blind Multi-Agent Arena Evaluations.

This dataset was generated using the DualBlind AI Benchmark Arena. In this setup, two independent frontier AI models engage in multi-turn double-blind dialogue to solve extreme-difficulty benchmark problems, verifying their peer's proofs, raising counter-examples, and reaching mathematical consensus.

Dataset Structure & Strict Quality Gate

Strict 100% Accuracy Quality Gate (Active): Only trials that achieved 100% automated verification against canonical ground-truth solutions are admitted into GlimmaryKarl/DualBlind. Incomplete trials, partial solutions (<100%), and unverified outputs are strictly blocked from entry.
  • —Total Arena Trials Evaluated: 3363
  • —Verified 100% Accurate Trials Admitted: 2992 (89% pass rate)
  • —Ineligible Trials Filtered Out (<100% or Refuted): 371
  • —Quality Gate Policy: Zero-tolerance for unverified or imperfect conjectures. Every single sample provides a mathematically and algorithmically verified canonical solution.

Subsets:

  1. 1.`sft_reasoning` (2962 samples): Supervised Fine-Tuning records containing high-density step-by-step chain-of-thought reasoning traces culminating in 100% verified ground-truth solutions.
  2. 2.`dpo_preferences` (2961 pairs): Direct Preference Optimization pairs (prompt, chosen, rejected) pairing the 100% verified consensus proof as chosen against flawed peer conjectures or refuted hypotheses as rejected.

Benchmark Suites Covered

  • —GPQA Diamond: 411 runs
  • —MMLU-Pro: 270 runs
  • —FrontierMath: 190 runs
  • —Game Theory: 426 runs
  • —ARC Challenge: 348 runs
  • —SWE-bench: 393 runs
  • —IFEval: 236 runs
  • —Humanity's Last Exam: 242 runs
  • —Formal Logic: 310 runs
  • —MATH / AIME: 143 runs
  • —General Benchmark: 23 runs

Models Represented

  • —meta-llama/llama-3.3-70b-instruct:free
  • —qwen/qwq-32b:free
  • —openrouter/free
  • —gemini-1.5-flash
  • —qwen/qwen-2.5-coder-32b-instruct:free
  • —gemini-2.5-flash
  • —qwen/qwen-2.5-72b-instruct:free
  • —deepseek/deepseek-r1:free
  • —microsoft/phi-3-mini-128k-instruct:free
  • —deepseek/deepseek-chat:free
  • —google/gemini-2.0-flash-exp:free
  • —deepseek/deepseek-r1-distill-llama-70b:free
  • —meta-llama/llama-3.1-8b-instruct:free
  • —mistralai/mistral-7b-instruct:free
  • —gemini-2.0-flash
  • —poolside/laguna-s-2.1:free
  • —poolside/laguna-xs-2.1:free
  • —nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free
  • —z-ai/glm-5.2:free
  • —nvidia/nemotron-3.5-lightning:free
  • —nvidia/nemotron-3-ultra-550b-a55b:free
  • —meta-llama/llama-3.2-3b-instruct:free
  • —mistralai/mistral-small-24b-instruct-2501:free
  • —cognitivecomputations/dolphin3.0-r1-mistral-24b:free
  • —gemini-3.7-flash
  • —google/gemma-2-9b-it:free
  • —nvidia/llama-3.1-nemotron-70b-instruct:free
  • —openai/gpt-3.5-turbo
  • —thedrummer/skyfall-36b-v2
  • —liquid/lfm-2.5-2.6b:free
  • —google/gemma-4-26b-a4b-it:free
  • —minimax/minimax-m3:free
  • —z-ai/glm-4.7-flash
  • —tencent/hy-mt2-30b-a3b
  • —openai/gpt-4o-mini
  • —meta-llama/llama-3.3-70b-instruct
  • —qwen/qwen-2.5-72b-instruct
  • —meta/muse-glimmer-30b
  • —google-gemini-3.7-flash
  • —claude-3-7-sonnet-20250219
  • —deepseek-r1
  • —o3-mini
  • —gpt-4.5-preview
  • —ox-alpha
  • —nvidia/nemotron-3-nano
  • —amazon/nova-lite-v1
  • —meta-llama/llama-3.1-70b-instruct
  • —deepseek-deepseek-v3
  • —anthropic-claude-3-haiku
  • —gpt-4o

Usage with Hugging Face Datasets

python
from datasets import load_dataset

# Load SFT Reasoning split
sft_ds = load_dataset("GlimmaryKarl/DualBlind", "sft_reasoning", split="train")
print(sft_ds[0]["messages"])

# Load DPO Preference split
dpo_ds = load_dataset("GlimmaryKarl/DualBlind", "dpo_preferences", split="train")
print(dpo_ds[0]["chosen"])
print(dpo_ds[0]["rejected"])

100% Free Fine-Tuning Guide (Unsloth on Google Colab T4)

You can fine-tune Llama-3.1-8B-Instruct or Qwen-2.5-7B in ~25 minutes on a free Google Colab T4 GPU:

bash
pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
pip install --no-deps trl peft accelerate bitsandbytes
python
from unsloth import FastLanguageModel
from trl import SFTTrainer
from transformers import TrainingArguments
from datasets import load_dataset

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/Llama-3.1-8B-Instruct",
    max_seq_length=2048,
    load_in_4bit=True,
)

model = FastLanguageModel.get_peft_model(
    model,
    r=16,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
    lora_alpha=16,
    lora_dropout=0,
    bias="none",
)

dataset = load_dataset("GlimmaryKarl/DualBlind", "sft_reasoning", split="train")

trainer = SFTTrainer(
    model=model,
    train_dataset=dataset,
    dataset_text_field="text",
    max_seq_length=2048,
    tokenizer=tokenizer,
    args=TrainingArguments(
        per_device_train_batch_size=2,
        gradient_accumulation_steps=4,
        warmup_steps=5,
        max_steps=60,
        learning_rate=2e-4,
        fp16=True,
        logging_steps=1,
        output_dir="outputs",
    ),
)
trainer.train()

License

Apache 2.0. Free for commercial, research, and educational use.