File size: 3,146 Bytes
f8b57d0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
---
license: cc-by-4.0
task_categories:
  - question-answering
  - text-generation
language:
  - en
tags:
  - stem
  - evaluation
  - benchmark
  - mmlu
  - gpqa
  - gsm8k
  - math
size_categories:
  - 1K<n<10K
---

# nanochat-npu-stem-eval

Pre-processed STEM evaluation data for [nanochat-npu](https://github.com/liujin99/nanochat-npu), adapted from karpathy/nanochat for Huawei 910B3 NPU.

## Tasks

| Task | Type | Shot | Source | Examples | Description |
|------|------|------|--------|----------|-------------|
| `gpqa_diamond` | multiple_choice | 0-shot | Idavidrein/gpqa | 198 | Graduate-level science QA (Diamond subset) |
| `gsm8k_cot` | generation | 8-shot | openai/gsm8k | 1319 | Grade school math word problems (CoT) |
| `math_cot` | generation | 4-shot | HuggingFaceH4/MATH-500 | 500 | Competition mathematics (CoT) |
| `mmlu_zeroshot` | multiple_choice | 0-shot | cais/mmlu | 14042 | MMLU full 57 subjects |
| `mmlu_stem` | multiple_choice | 0-shot | cais/mmlu | 3545 | MMLU STEM subset (22 subjects) |

## MMLU STEM Subjects (22)

abstract_algebra, anatomy, astronomy, college_biology, college_chemistry, college_computer_science, college_mathematics, college_physics, computer_security, conceptual_physics, electrical_engineering, elementary_mathematics, formal_logic, high_school_biology, high_school_chemistry, high_school_computer_science, high_school_mathematics, high_school_physics, high_school_statistics, machine_learning, medical_genetics, virology

## Data Format

### Multiple Choice Tasks (gpqa_diamond, mmlu, mmlu_stem)

```json
{
  "query": "Question text\nChoices:\n(A) choice0\n(B) choice1\n(C) choice2\n(D) choice3",
  "choices": ["choice0", "choice1", "choice2", "choice3"],
  "gold": 2
}
```

`gold` is the index (0-3) of the correct answer. GPQA choices are shuffled with seed=42.

### Generation Tasks (gsm8k, math500)

```json
{
  "question": "Problem text",
  "answer": "Full solution text",
  "gold_answer": "42"
}
```

`gold_answer` is the extracted final answer (only in MATH-500). For GSM8K, the answer contains `#### N` which is extracted at evaluation time.

## Usage

### Download as evaluation bundle (recommended)

The `eval_stem.zip` contains all data in the exact layout expected by nanochat evaluation scripts. It is downloaded automatically on first evaluation run.

### Load individual tasks with datasets library

```python
from datasets import load_dataset

# Load MMLU STEM subset
ds = load_dataset("liujin99/nanochat-npu-stem-eval", "mmlu_stem")
# Load GSM8K
ds = load_dataset("liujin99/nanochat-npu-stem-eval", "gsm8k")
```

## Source Datasets and Licenses

| Dataset | Source | License |
|---------|--------|---------|
| GPQA Diamond | [Idavidrein/gpqa](https://huggingface.co/datasets/Idavidrein/gpqa) | CC-BY-4.0 |
| GSM8K | [openai/gsm8k](https://huggingface.co/datasets/openai/gsm8k) | MIT |
| MATH-500 | [HuggingFaceH4/MATH-500](https://huggingface.co/datasets/HuggingFaceH4/MATH-500) | MIT |
| MMLU | [cais/mmlu](https://huggingface.co/datasets/cais/mmlu) | MIT |

All data is redistributed with attribution under the original licenses. See `LICENSES.txt` in the zip for details.