NickIBrody/x86-asm-instructions-23k

---
license: other
task_categories:
- text-generation
language:
- en
tags:
- code
- assembly
- x86
- x86-64
- instruction-tuning
- code-generation
size_categories:
- 10K<n<100K
pretty_name: x86 ASM Instructions - 23K
---

# x86 ASM Instructions - 23K

Instruction-tuning dataset of x86 and x86_64 assembly snippets paired with short natural-language instructions or comments.

## Dataset Summary

| Field | Value |
|---|---|
| Total rows | 23,104 |
| Split(s) | train |
| Format | JSONL (`instruction`, `output`, `system`) |
| Instruction average length | 47.0 characters |
| Instruction max length | 413 characters |
| Output average length | 514.25 characters |
| Output max length | 15,805 characters |
| System prompt values | 1 |

## Format

Each example contains:

```json
{
"instruction": "Recursive factorial. cdecl calling convention.",
"output": "factorial_rec_cdecl:\n enter 0, 0\n ...",
"system": "As an x86/x86_64 assembly expert, you are capable of writing optimized low-level code from specifications."
}
```

## Recommended Use

- supervised fine-tuning for x86 or x86_64 assembly generation
- low-level code completion experiments
- instruction-to-assembly research with additional downstream filtering

## Important Limitations

- the dataset currently exposes only a `train` split
- instructions vary in quality and appear to include extracted comments, partial descriptions, and terse labels rather than consistently written prompts
- the dataset appears to mix multiple domains including general assembly routines, DOS-era utilities, boot-related code, and potentially malware-adjacent samples
- examples are standalone snippets and often rely on missing surrounding context such as memory layout, calling convention, macro definitions, or target platform
- the `system` field is constant across rows and may be better injected by the training template instead of consumed as sample data

## Safety Note

This dataset appears to include low-level code related to file infection, boot manipulation, destructive disk access, or other malware-adjacent behavior in at least some samples. Downstream users should review and filter the corpus before production use, safety-sensitive training, or redistribution.

## Provenance And Licensing

This release does not currently document per-example provenance fields such as repository, source file, commit, or source license.

Because of that, this dataset should not be treated as a clean permissively licensed corpus without further legal and provenance review.

## Missing Metadata

The current release still lacks:

- repository / file / commit provenance
- per-example source license metadata
- domain labels such as `bootloader`, `dos`, `utility`, `malware`, or `reversing`
- validation and test splits
- contamination analysis against downstream benchmarks

## Suggested Next Steps

To make this dataset more useful, the next revision should add:

1. deterministic `validation` and `test` splits
2. provenance fields for each sample
3. filtering or explicit labeling for malware-adjacent content
4. instruction-quality filtering for short or fragmentary prompts