samscrack's picture
Initial release: 23,471 quality-filtered + deduped Solidity files (~86M tokens)
23c0b2f verified
---
license: cc-by-4.0
language:
- en
- code
tags:
- solidity
- smart-contracts
- blockchain
- ethereum
- code
- continued-pretraining
- quality-filtered
size_categories:
- 10K<n<100K
task_categories:
- text-generation
- fill-mask
pretty_name: Solidity CPT Top-10% Quality-Filtered Corpus
configs:
- config_name: default
data_files:
- split: train
path: top10.parquet
---
# Solidity CPT Top-10% Quality-Filtered Corpus
A curated, deduplicated corpus of **23,471 modern Solidity source files (~86M tokens)** intended for continued-pretraining (CPT) of code LLMs on smart-contract code.
It's the top 10% slice (by composite quality score) of a larger raw corpus that combined:
- [`ASSERT-KTH/DISL`](https://huggingface.co/datasets/ASSERT-KTH/DISL) — 514 k unique deployed Solidity files, deduped at file level
- 30 hand-picked GitHub blue-chip protocols (OpenZeppelin, Uniswap v2/v3/v4, Aave v3, Compound, Morpho, EigenLayer, Pendle, Solady, Seaport, LayerZero, ENS, Optimism, Arbitrum, Polygon L1, etc.)
After scoring all 234 877 raw rows on a 0–115 quality scale (see *Quality filter* below) and keeping the top 10 %, the result is **23 487 source files**. After cross-source SHA-256 deduplication of whitespace-normalised text, **23 471 unique files** remain. This is what's published here.
## At a glance
| | |
|---|---|
| Rows | 23,471 |
| Total characters | 342,858,470 |
| Approx. tokens (Qwen3.6 tokenizer, ~4 chars/tok) | **~86M** |
| Average chars per row | 14,607 |
| Source: `disl` (Etherscan-verified) | 20,975 (89 %) |
| Source: `github/<protocol>` (audited blue-chips) | 2,496 (11 %) |
| File format | parquet (snappy) + jsonl |
## Schema
Each row is one Solidity source file:
| Field | Type | Description |
|---|---|---|
| `text` | string | Raw Solidity source code |
| `source` | string | `disl` (Etherscan-verified contract) or `github/<repo-slug>` |
| `address` | string | Ethereum mainnet address (DISL only) |
| `name` | string | Contract name |
| `compiler` | string | Solc version e.g. `v0.8.20+commit.a1b2c3d4` (DISL only) |
| `license` | string | License from Etherscan (often empty) |
| `path` | string | Path inside the GitHub repo (GitHub rows only) |
| `n_chars` | int | Length of `text` in characters |
## Quality filter (composite 0–115 score)
Every raw row was scored independently on each signal; rows scoring **≥ 55 (top-10% threshold)** were kept.
| Signal | Points |
|---|---:|
| Pragma 0.8.20+ (modern compiler) | +30 |
| Pragma 0.8.13–19 | +20 |
| Pragma 0.8.x (any) | +10 |
| Has SPDX license header | +10 |
| Has `@notice` / `@dev` NatSpec | +10 |
| Has `@param` / `@return` NatSpec | +10 |
| Comment density 5–30 % | +10 |
| Uses custom errors (`error X();`) | +5 |
| Uses `unchecked { ... }` blocks | +5 |
| No SafeMath import (avoids old patterns) | +5 |
| Size 500–8000 chars | +10 |
| Source: GitHub blue-chip | +20 |
**Pre-scoring filters** (applied to all sources before scoring):
- Pragma must be 0.8.x (0.4–0.7 dropped — incompatible with `forge-std/Test.sol >=0.8.13`)
- Not flagged as `proxy=True` in DISL (drops EIP-1167/1967 minimal proxies)
- Has at least one `contract X` declaration (drops interface-only / library-only files)
- 200 ≤ length ≤ 50000 chars (drops stubs and mega-flatteners)
- For GitHub: vendored `lib/openzeppelin-contracts`, `lib/forge-std`, `out/`, `cache/`, `typechain-types/` stripped
**Post-quality dedup**: SHA-256 of the whitespace-collapsed text. Caught 16 cross-source duplicates.
## Score distribution of the top-10% slice
The 23 471 surviving rows have scores in [55, 115]. The mode falls in the 55–69 band (typical "modern, idiomatic, audited" Solidity). The long tail (90+) skews heavily toward GitHub blue-chips and DISL contracts that cite their license + use NatSpec extensively.
## Suggested use
This corpus is sized for **adapter-based CPT** (LoRA / DoRA) on a 7B–30B+ code LM. Per [Biderman et al. 2024 ("LoRA Learns Less and Forgets Less")](https://arxiv.org/abs/2405.09673), useful adapter-CPT plateaus around 200M–1B tokens — this 86M corpus is on the lower end of that range and works well as a fast first-iteration validation slice.
For larger adapter-CPT runs, the same pipeline produces:
- **top30** tier (~70 k rows, ~229M tokens)
- **top60** tier (~141 k rows, ~458M tokens)
(Not yet published. Open an issue if you'd like them released.)
## Loading
### `datasets` library (recommended)
```python
from datasets import load_dataset
ds = load_dataset("samscrack/solidity-cpt-top10-quality", split="train")
print(ds) # ~23 471 rows
print(ds[0]["text"][:500])
```
### Direct parquet
```python
import pyarrow.parquet as pq
table = pq.read_table("top10.parquet")
df = table.to_pandas()
print(df.head())
```
### Streaming (for low-RAM environments)
```python
ds = load_dataset(
"samscrack/solidity-cpt-top10-quality",
split="train",
streaming=True,
)
for row in ds.take(3):
print(row["source"], row["name"], row["compiler"])
```
## Suggested CPT recipe (Qwen3.6-27B reference)
This corpus was built for the [Qwopus3.6-27B-solidity-cpt-stageA](https://huggingface.co/samscrack/Qwopus3.6-27B-solidity-cpt-stageA) pipeline. The recipe used:
- LoRA r=64, α=64, target modules: `q_proj` `k_proj` `v_proj` `o_proj` `gate_proj` `up_proj` `down_proj` `out_proj`
- Megatron-style packing into 8192-token sequences with EOS separators
- LR 5e-5 (cosine), bf16 + 4-bit base (QLoRA), 1 epoch
- Tokenized with the Qwen3.6 tokenizer (vocab 152 k)
A Stage-A CPT over a 1500-sequence (~12M token) subset of this corpus showed loss drop from a warmup peak of 0.67 to a stable 0.36–0.41 plateau in ~40 steps — clear evidence the data is high-signal for adapter-CPT.
## License & legal notes
This dataset is released under **CC BY 4.0**, inheriting the most-restrictive upstream license (DISL is CC BY 4.0).
**Per-row provenance varies**:
- DISL rows: derived from Etherscan public-verified contracts. The dataset (and DISL itself) are CC BY 4.0; **the underlying contract source has whatever license the deployer chose** — often empty / "unspecified" / all-rights-reserved by default. Suitable for research and adapter-style CPT, not necessarily safe for redistribution-as-source.
- GitHub blue-chip rows: each repo has its own SPDX header. Most are MIT or Apache-2.0; some Uniswap pieces are BUSL/GPL. Inspect the `path` field and the upstream repo if license matters for your use case.
If you need stronger commercial-use guarantees, restrict to rows where:
- `source` starts with `github/`, AND
- `text` contains an explicit MIT or Apache-2.0 SPDX line near the top.
## Citation
```bibtex
@misc{solidity_cpt_top10_2026,
title = {Solidity CPT Top-10\% Quality-Filtered Corpus},
author = {samscrack},
year = {2026},
url = {https://huggingface.co/datasets/samscrack/solidity-cpt-top10-quality}
}
```
If you use this data, please also cite the upstream sources:
```bibtex
@misc{disl_2024,
title = {DISL: Fueling Research with A Large Dataset of Solidity Smart Contracts},
author = {Storhaug, Andreas et al.},
year = {2024},
url = {https://arxiv.org/abs/2403.16861}
}
```
## Acknowledgements
- [ASSERT-KTH](https://huggingface.co/ASSERT-KTH) for releasing DISL
- The OpenZeppelin, Uniswap, Aave, Compound, Morpho, EigenLayer, Pendle, Solady, Seaport, LayerZero, ENS, Optimism, Arbitrum, and Polygon teams for keeping their protocol code public and auditable
- The Foundry team for forge-std (used as the `>=0.8.13` floor for pragma filtering)
## Changelog
- **2026-05-03** — initial release; 23 471 rows, ~86M tokens, top 10% by composite quality score