Aswina1636/mimo-rl-curated-code

MiMo-RL Curated (code)

A cleaned, flattened and labeled copy of the code subset of XiaomiMiMo/MiMo-V2.6-RL-oss, with extra columns that make it easier to filter, sample and explore. It is an independent project and is not affiliated with or endorsed by Xiaomi MiMo. All task texts and test patches come from the original dataset, unchanged.

• full: 2,698 tasks. lite: 500 tasks (25 per topic cluster), small enough for a laptop.
• Only the code subset is covered (not webdev, music, cyber or general).
• License: Apache-2.0, the same as the original dataset. Please also read the original dataset card for any extra terms, and note that task texts describe third-party open-source projects.

What was added

Added	What it is
Flat columns	The original nested prompt, reward_model and extra_info (with a JSON string inside) are unpacked into plain columns.
Profile columns	prompt length, code block / traceback flags, Chinese-text flag, test files, patch size.
prog_language	Language of the test files (rule-based).
difficulty	An estimated easy / medium / hard group.
issue_type	Predicted bug or feature, with a confidence score.
cluster_id, cluster_name	20 automatic topic groups.
near_duplicate	Flags very similar tasks.

Removed because they carry no information: data_source, ability, agent_name (the same value in every row), the empty reward_model.ground_truth, and the duplicate problem_statement (identical to prompt).

Quick start

from datasets import load_dataset

lite = load_dataset("Aswina1636/mimo-rl-curated-code", "lite", split="train")
easy_english = lite.filter(lambda r: r["difficulty"] == "easy" and not r["has_chinese"])
sure_bugs = lite.filter(lambda r: r["issue_type"] == "bug" and r["issue_type_confident"])

embeddings/mpnet_embeddings.parquet holds one all-mpnet-base-v2 vector per task (key: instance_id) for similarity search.

Columns

Column	Description
instance_id	Task id from the original dataset (unique key).
orig_index	Row index in the original dataset's code subset.
prompt	The task text (a GitHub-issue style description). Unchanged from the original.
test_patch	Diff of the test files that verify the task. Unchanged from the original.
docker_image	Image name given in the original dataset. The images themselves are NOT included here.
verifier_timeout_sec	Verifier timeout from the original dataset (1800 for every task).
prog_language	Programming language, guessed from the real test files in test_patch (rule-based; Other = no known extension).
test_files	List of real test files touched by test_patch (harness files excluded, see Changelog).
n_test_files	Number of real test files in test_patch (harness files excluded).
patch_lines	Added plus removed lines in test_patch, harness files excluded.
prompt_words	Words in the prompt.
prompt_chars	Characters in the prompt.
has_code_block	Prompt contains a markdown code fence.
has_traceback	Prompt contains error/traceback words (rough text flag).
has_chinese	Prompt contains CJK characters (flags Chinese and some Japanese text).
difficulty	easy / medium / hard: an ESTIMATE in three equal groups (see 'How the columns were made').
difficulty_score	0-1 score behind the difficulty groups.
issue_type	Predicted: bug or feature (see limits).
issue_type_confidence	Classifier probability (0.5-1.0) of the predicted issue_type.
issue_type_confident	True when issue_type_confidence >= 0.70.
cluster_id	Topic cluster number (0-19).
cluster_name	Short human-written name for the cluster. Names are approximate.
near_duplicate	True when the closest other task has similarity above 0.90.
nearest_instance_id	Closest other task (by all-MiniLM-L6-v2 similarity).
nearest_similarity	Cosine similarity to that closest task.

How the columns were made

• prog_language: first real test file in test_patch whose extension matches a known language. It is the language of the tests; a project in one language can have shell-script tests (Shell is 72 tasks).
• difficulty: an estimate, not measured model performance. Score = 40% test-patch size + 30% number of real test files + 30% prompt length, each turned into a percentile rank, then cut into three equal groups. "Hard" means "larger than most other tasks here". A short task can still be hard.
• issue_type: logistic regression on all-mpnet-base-v2 embeddings + TF-IDF words + a few text flags, trained on about 350 labeled tasks (mostly labeled by an LLM, plus some manual labels and corrections). Only two classes are predicted; the roughly 9% of tasks that are neither a bug nor a feature (refactors, docs, performance, unclear) are predicted as bug or feature anyway.
• cluster_id / cluster_name: KMeans (k = 20) on all-MiniLM-L6-v2 embeddings. Names were written by reading samples and are approximate. Some clusters group by writing style (many prompts start "I want...") as well as by topic.
• near_duplicate / nearest_*: cosine similarity on all-MiniLM-L6-v2 embeddings (threshold 0.90). The published embeddings use a different model (all-mpnet-base-v2), so neighbours found with them can differ.
• has_chinese: any CJK character in the prompt.

Accuracy of issue_type (read this before relying on it)

Measure	Result
Bug vs feature, all tasks that were clearly one of the two (n = 228, cross-validated)	87.4%
Only predictions with confidence >= 0.70 (81.1% of those tasks)	93.9%
Only predictions with confidence >= 0.80 (71.2% of those tasks)	97.1%

• Truth for these numbers is the original label of each test task, so later label corrections cannot raise the score.
• The test set is small (228 tasks): expect about plus or minus 4-5 points.
• No independent human validation has been done yet. The accuracy numbers above measure agreement with labels that were mostly produced by an LLM, so they may overstate agreement with human judgment.
• Predictions are less reliable for Chinese-language prompts (average confidence about 0.73 vs 0.88 for the rest).
• For analysis, filter on issue_type_confident when you need labels you can trust.

Other limits

• difficulty and cluster_name are heuristics, not ground truth.
• Do not split this data randomly into train and evaluation sets. Tasks from the same project can be close neighbours (8 pairs score above 0.80 similarity) and can share context. Repository names are not stored in the dataset, so a split by repository needs them to be recovered from the task text first. lite is balanced by cluster and is not meant as a train/eval split.
• Languages are unbalanced: Python 1178, Go 721, JavaScript 388, TypeScript 166, Shell 72, Other 63, Ruby 25, PHP 23, Java 22, C++ 13, C 7, Rust 7, Scala 6, Julia 4, Kotlin 3.
• lite is balanced on purpose (25 per cluster), so it is not a random sample: it has more Chinese prompts (5.6% vs 3.4%) and a lower share of confident labels than the full set.
• The Docker images and verifiers of the original environments are not included. To run the tasks, use the original dataset.

Clusters (full set)

Cluster	Tasks
Python library features	192
CLI & config	184
State, jobs & lifecycle	177
Queries, filters & metrics	162
Data science (numpy, pandas...)	156
Config, paths & metadata	154
Python tooling	152
HTTP & APIs	150
Text, markup & parsing	145
Infra & ops bug reports	144
Data formats & library APIs	142
JavaScript libraries & tooling	132
Cloud-native & infra (Go)	129
Auth, users & permissions	123
Retries, timeouts & connections	123
SQL & databases	119
Frontend & UI	108
Go library features	79
Git, versions & releases	65
Chinese & non-English prompts	62

Changelog

• v1.1: test_files, n_test_files, patch_lines, prog_language and difficulty now ignore three helper files that the original environments add to test patches (mimo_test_command.sh in every task, mimo_build_env.tar.gz.b64 in about 31%, test_commands.json in about 22%). Before, every task counted at least one extra test file, 63 tasks were labeled Shell only because of the helper script, and about 21% of tasks sat in a different difficulty group. The lite split was rebuilt.
• v1.0: first release.

Credit

Task data: XiaomiMiMo/MiMo-V2.6-RL-oss (Apache-2.0). Added columns and tooling: this project, https://github.com/ASWINa1636/MiMo-RL-Curated.