---
language:
- en
license: apache-2.0
task_categories:
- text-generation
tags:
- perfectblend
- synthetic
- reasoning
- alignment
- multi-turn
size_categories:
- 1M<n<10M
configs:
- config_name: default
  data_files:
  - split: train
    path: data/*.parquet
---

# PerfectBlend Unified Multi-Model Collation

A unified collation of 44 open PerfectBlend regenerated datasets across diverse reasoning and non-reasoning language models, aligned onto a single wide schema by exact model input prompts.

## Dataset Overview

- **Total Rows (Distinct Prompts)**: 6,866,154
- **Total Columns**: 46 (1 `prompt` column + 45 dataset response columns)
- **Source Repositories Collated**: 44 repositories from the PerfectBlend regeneration initiative
- **Format**: 64 Snappy-compressed Parquet shards (`train-00000-of-00064.parquet` .. `train-00063-of-00064.parquet`)
- **Total Dataset Size**: ~90 GB compressed

## Schema & Field Specifications

### `prompt`
- **Type**: String (canonical JSON array)
- **Format**: `[{"role": "user"|"system"|"assistant", "content": "..."}]`
- **Description**: Preserves the complete model input context without loss, including system prompts, multi-turn conversational history, and the final user instruction. Generated model completions are excluded from this column.

### Dataset Response Columns
Each response column corresponds to one source dataset repository, sanitized using the rule `[owner]-[repo]` with special characters replaced by hyphens:
- **Reasoning Models**: Responses are standardized so reasoning traces appear inside `<think>...</think>` tags, followed by the final answer:
  ```text
  <think>
  step-by-step reasoning
  </think>

  final answer
  ```
- **Non-Reasoning Models**: Contains the direct generated completion without `<think>` delimiters.
- **Missing Responses**: Represented as `null` where a model dataset does not contain an evaluation or completion for that specific prompt.

## Column Distribution & Non-Null Coverage

| Column Name | Non-Null Completions | Coverage % |
| :--- | :--- | :--- |
| `3podi-perfectblend-qwen4B-regenerated` | 854,366 | 12.44% |
| `AQ-MedAI-Kimi-K2-Instruct-open-perfectblend-regenerate` | 889,954 | 12.96% |
| `AQ-MedAI-Ling-flash-2-0-open-perfectblend-regenerate` | 890,640 | 12.97% |
| `Cornmonster-kimi-k25-perfectblend` | 390,143 | 5.68% |
| `Dogacel-open-perfectblend-kimi-linear-regen` | 882,821 | 12.86% |
| `JingweiSong-perfectblend-qwen3-8b-regen-no-thinking` | 890,660 | 12.97% |
| `Rano23-qwen3-14b-perfectblend-regen` | 855,439 | 12.46% |
| `TIE-Pilot-qwen3-4b-perfectblend-deepspec-rollout` | 890,660 | 12.97% |
| `alice1001-open-perfectblend-qwen3-4b-regen` | 849,760 | 12.38% |
| `alice1001-open-perfectblend-qwen3-8-27b-regen` | 890,455 | 12.97% |
| `baseten-admin-gpt-oss120b-generated-perfectblend` | 890,514 | 12.97% |
| `baseten-admin-perfectblend-Qwen3-235B-A22B-Instruct-2507-FP8-generated` | 890,658 | 12.97% |
| `cm00cm-kimi-k2-7-open-perfectblend-regenerated` | 850,562 | 12.39% |
| `frankleeeee-PerfectBlend-Regenerated-Llama-3-1-8B-Instruct` | 889,502 | 12.95% |
| `frankleeeee-PerfectBlend-Regenerated-Llama-3-3-70B-Instruct` | 890,083 | 12.96% |
| `frankleeeee-PerfectBlend-Regenerated-Llama-4-Maverick-17B-128E-Instruct` | 890,643 | 12.97% |
| `frankleeeee-PerfectBlend-Regenerated-Llama-4-Scout-17B-16E-Instruct` | 884,101 | 12.88% |
| `heiheiha798-perfectblend-regen-qwen3-8b-non-thinking` | 890,167 | 12.96% |
| `jihwan1205-perfectblend-qwen3-8b-regen` | 888,221 | 12.94% |
| `kimyuji-perfectblend-qwen3-8b-full-shuffled` | 878,018 | 12.79% |
| `langq1225-perfectblend-qwen3-4b-regen` | 889,485 | 12.95% |
| `linYD0718-open-perfectblend-qwen3-4b-nonthinking` | 855,134 | 12.45% |
| `lukeysong-qwen3-235-regen-perfect-blend` | 886,822 | 12.92% |
| `mgoin-open-perfectblend-glm5-2-regen` | 890,328 | 12.97% |
| `mocoV3-Open-PerfectBlend-qwen3-5-4B-regenerate` | 613,192 | 8.93% |
| `rasyosef-open-perfectblend-gemma-4-E4B-it` | 26,526 | 0.39% |
| `rasyosef-open-perfectblend-phi-4-mini-instruct` | 106,561 | 1.55% |
| `rasyosef-open-perfectblend-qwen-3-5-2b` | 98,675 | 1.44% |
| `rasyosef-open-perfectblend-qwen3-1-7b` | 103,824 | 1.51% |
| `rasyosef-open-perfectblend-qwen3-1-7b-nothink` | 51,850 | 0.76% |
| `shanjiaz-glm-5-3-flash-open-perfectblend-regen-25pct` | 82,158 | 1.20% |
| `shanjiaz-qwen3-5-4b-perfectblend-regen` | 1,196,729 | 17.43% |
| `shuaills-perfect-blend-gptoss-20B` | 87,839 | 1.28% |
| `sunnysetia-GLM-5-2-PerfectBlend-Regen-Sharded` | 682,572 | 9.94% |
| `susun-123-perfectblend-qwen3-4b-regen` | 663,412 | 9.66% |
| `xf15-ssd-perfectblend-qwen3-30b-a3b-regen` | 855,433 | 12.46% |
| `xf15-ssd-perfectblend-qwen3-4b-regen` | 855,433 | 12.46% |
| `xpzhao11-open-perfectblend-qwen38-flash-next-regen` | 855,462 | 12.46% |
| `yosefw-open-perfectblend-gemma-4-E2B-it` | 99,468 | 1.45% |
| `yosefw-open-perfectblend-llama-3-2-3b-instruct` | 115,251 | 1.68% |
| `yosefw-open-perfectblend-qwen-3-1-7b` | 1,023 | 0.01% |
| `yosefw-open-perfectblend-qwen-3-5-2b` | 53,316 | 0.78% |
| `yuekai-perfectblend-qwen3-omni-sglang` | 850,109 | 12.38% |
| `zhuyksir-perfect-blend-gptoss-20B-1M` | 833,943 | 12.15% |
| `zhuyksir-perfectblend-Qwen3-235B-A22B-Instruct-2507-FP8-generated` | 548,472 | 7.99% |

## Quick Start & Usage

### Hugging Face Datasets

```python
from datasets import load_dataset

# Stream the dataset without downloading all shards at once
dataset = load_dataset("MaxDevv/perfectblend-unified", split="train", streaming=True)

for example in dataset.take(2):
    print("Prompt:", example["prompt"][:200])
    # Inspect specific model outputs
    for col, val in example.items():
        if col != "prompt" and val is not None:
            print(f"  {col}: {val[:150]}...")
```

### DuckDB

```sql
SELECT prompt, "3podi-perfectblend-qwen4B-regenerated"
FROM read_parquet('hf://datasets/MaxDevv/perfectblend-unified/data/*.parquet')
WHERE "3podi-perfectblend-qwen4B-regenerated" IS NOT NULL
LIMIT 5;
```

### Polars

```python
import polars as pl

# Scan dataset with predicate and projection pushdown
lf = pl.scan_parquet("data/*.parquet")
sample = lf.select(["prompt", "rasyosef-open-perfectblend-qwen-3-5-2b"]).drop_nulls().limit(5).collect()
print(sample)
```

## Data Sample

```json
{
  "prompt": "[{\"content\": \"If John books a hotel room for 3 nights at a rate of $250 per night and receives a $100 discount, what is the total amount he needs to pay?\", \"role\": \"user\"}]",
  "3podi-perfectblend-qwen4B-regenerated": "John is booking a hotel room for **3 nights** at a rate of **$250 per night**, and he receives a **$100 discount**.\n\n### Step 1: Calculate the total cost without the discount...",
  "Cornmonster-kimi-k25-perfectblend": "<think>\nThe user wants to calculate the total amount John needs to pay for a hotel room.\n1. Cost per night = $250\n2. Number of nights = 3\n3. Total cost before discount = 3 * 250 = $750\n4. Discount = $100\n5. Total to pay = 750 - 100 = $650\n</think>\n\nTo find the total amount John needs to pay, we can break down the calculation into simple steps:\n\n1. **Calculate the cost before discount:**\n   $$\\text{Total cost} = 3 \\text{ nights} \\times \\$250/\\text{night} = \\$750$$\n\n2. **Apply the discount:**\n   $$\\text{Final amount} = \\$750 - \\$100 = \\$650$$\n\nJohn needs to pay **$650**."
}
```

## Extraction & Normalization Reproducibility
- Extracted and normalized from source repositories in `perfectblend_datasets.txt`.
- Duplicate prompts deduplicated per dataset; cross-model responses aligned on canonical prompt JSON.