eigentom/minicpm5-sft-swe-validation-200

MiniCPM5 SFT SWE Validation 200

A fixed, public task index for small-scale software-engineering evaluation. It contains two configurations, verified and pro, each with exactly 100 distinct tasks: 70 that the historical MiniCPM5-2B SFT baseline resolved and 30 that it did not resolve. Every row has benchmark, instance_id, task_id, sft_resolved, repo, and base_commit.

The task statements, repository contents, reference patches, and tests are not copied into this dataset. Join instance_id to the original benchmark to run the tasks:

• verified: princeton-nlp/SWE-bench_Verified, test split (500 tasks).
• pro: ScaleAI/SWE-bench_Pro, v1 config, test split (731 tasks). Do not use the later default v2 config to materialize this selection.

Example:

from datasets import load_dataset

selection = load_dataset("eigentom/minicpm5-sft-swe-validation-200", "verified", split="validation")
source = load_dataset("princeton-nlp/SWE-bench_Verified", split="test")
by_id = {row["instance_id"]: row for row in source}
tasks = [by_id[row["instance_id"]] for row in selection]

For Pro, use load_dataset("ScaleAI/SWE-bench_Pro", "v1", split="test") and join the same way.

Selection and provenance

The baseline is the completed sft-verified-pro evaluation, run ID 09d1b154381c, with manifest SHA-256 78588b13c8d53a4f3f26719959d0bf31e7a1ed76335efc12834e429d1f1cd548. Its protocol records model ID openbmb/MiniCPM5-2B, revision abe115e887989b14f05e64a3b260648329324c3f, and one attempt per task. sft_resolved is the boolean result.resolved from tasks whose phase was complete. review and unfinished tasks were excluded. The eligible pools were 187 resolved and 290 unresolved Verified tasks, and 111 resolved and 560 unresolved Pro tasks.

Within each benchmark/outcome stratum, rows are ranked lexicographically by SHA256(seed + NUL + benchmark + NUL + int(resolved) + NUL + instance_id) and the first 70 resolved or 30 unresolved are selected. The seed is minicpm5-sft-swe-validation-200-v1. Final row order is separately hash-shuffled. manifest.json records the source asset and output hashes; build.py reproduces the files from the original cpudev run.

This set is deliberately balanced on the baseline outcome. Its baseline accuracy is 70% by construction and is not an estimate of accuracy on the full benchmarks. It is intended for checkpoint comparison and debugging; report full-benchmark results separately. Do not train on this public validation selection if using it for an unbiased comparison.

The original benchmark datasets and underlying repositories retain their own licenses and attribution requirements. This dataset distributes only task identifiers, factual repository metadata, and the baseline outcome labels.