Datasets:
Tasks:
Text Generation
Formats:
json
Languages:
English
Size:
< 1K
Tags:
reflection-ai
build-with-beam
beam-1.0-501b-a23b
mcp
legacy-modernization
differential-verification
License:
File size: 3,330 Bytes
f7c1e5e | 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 | ---
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- reflection-ai
- build-with-beam
- beam-1.0-501b-a23b
- mcp
- legacy-modernization
- differential-verification
- code-benchmarks
pretty_name: Legacy2MCP-Bench (Real GitHub Repositories · 160 Differential Vectors)
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: test
path: legacy2mcp_bench.jsonl
---
# Legacy2MCP-Bench (v2 · 100% Real GitHub Ground-Truth Corpus)
**`Legacy2MCP-Bench`** is a verifiable 160-vector enterprise modernization benchmark built for the **Reflection AI × Hugging Face "Build with Beam"** competition. Every row measures whether an LLM can refactor a real, production multi-file GitHub Python repository into a **zero-dependency standalone Anthropic Model Context Protocol (`MCPServer` v2)** module that achieves **bit-exact parity** against the untouched original GitHub library.
## Cloned GitHub Repositories & Commits
| Repo ID | Public GitHub Repository | Git Commit | Domain & Compliance Regime | Raw Cloned Tokens (`Beam-1.0` BPE) | AST-Compacted Pack (`Beam-1.0` BPE) | Vectors | Single-Chunk RAG Parity | Whole-Repo + Self-Heal Parity |
|---|---|---|---|---:|---:|---:|---:|---:|
| `py_moneyed` | [`limist/py-moneyed`](https://github.com/limist/py-moneyed) | `4c90ed5` | ISO-4217 Treasury & Pro-Rata Allocation | 15,826 | **8,660** | 40 | 37/40 (`92.5%`) | **40/40 (`100.0%`)** |
| `price_parser` | [`scrapinghub/price-parser`](https://github.com/scrapinghub/price-parser) | `6718bfe` | Global eCommerce Price & Locale Extraction | 19,246 | **16,495** | 40 | 29/40 (`72.5%`) | **40/40 (`100.0%`)** |
| `schwifty` | [`mdomke/schwifty`](https://github.com/mdomke/schwifty) | `7462c6e` | ISO-13616 IBAN / BBAN / SWIFT-BIC | 2,400,664 | **1,882** | 40 | 24/40 (`60.0%`) | **40/40 (`100.0%`)** |
| `python_stdnum` | [`arthurdejong/python-stdnum`](https://github.com/arthurdejong/python-stdnum) | `006192e` | ISIN / CUSIP / SEDOL / FIGI / ABA / EIN / EU VAT | 639,812 | **5,745** | 40 | 24/40 (`60.0%`) | **40/40 (`100.0%`)** |
| **Total** | **4 Real GitHub Repos (447 files)** | — | **Enterprise Finance, Tax & Commerce** | **3,075,548** | **32,782 (`98.9%` saved)** | **160** | **114/160 (`71.25%`)** | **160/160 (`100.00%`)** |
## Schema (`legacy2mcp_bench.jsonl`)
Each of the 160 JSONL rows contains:
- `benchmark_id`: Unique vector ID (`PY_MONEYED-01` .. `PYTHON_STDNUM-40`)
- `github_repo`, `github_url`, `git_commit_sha`: Exact upstream GitHub provenance
- `mcp_tool_name`, `category`, `label`, `input_arguments`: Structured MCP tool invocation payload
- `ground_truth_output`: Exact JSON output produced by executing the untouched GitHub library
- `synthesized_mcp_output`: Exact JSON output produced by executing the standalone LLM-synthesized `MCPServer` module (zero legacy library imports enforced via Python AST audit)
- `whole_repo_1m_exact_match` (`bool`), `single_chunk_rag_exact_match` (`bool`)
- `whole_repo_openrouter_request_ids`, `single_chunk_openrouter_request_id`: Verifiable OpenRouter generation IDs (`gen-...`)
- `official_beam_bpe_packed_tokens`: Exact token count measured with `reflection/Beam-1.0-501B-A23B` `tokenizer.json` (`vocab_size=151,681`)
- `parity_certificate_sha256`: Cryptographic verification digest
|