Clarify dataset curation, annotations, and usage

Document KodCode annotation provenance and statistics, distinguish upstream TACO/LeetCode tags and independent experiment splits, and update release file checksums.

CHANGELOG.md

@@ -1,5 +1,12 @@
 # Changes
 
+## Documentation update — 2026-09-09
+
+- Clarified the research purpose and our curation, annotation, and ablation work.
+- Added recorded KodCode annotation statistics, field definitions, and an example.
+- Distinguished upstream tags from our API annotations and documented split limits.
+- No changes to data, labels, splits, experimental memberships, or reward code.
+
 ## 1.0.0 — 2026-09-07
 
 - Four Parquet configurations with train and validation splits.

NOTICE.md

@@ -8,7 +8,13 @@
 - [LeetCodeDataset](https://arxiv.org/abs/2504.14655), 2025.
   Source: [newfacade/LeetCodeDataset](https://huggingface.co/datasets/newfacade/LeetCodeDataset).
 
-This release packages filtered project snapshots, preserves reference answers
-and RL tests, carries category labels, and adds a TACO SFT holdout.
-Upstream authors retain attribution. Exact revisions and file identities are
-recorded under metadata/. TACO's own source acknowledgments also apply.
+Our contributions are data curation, LLM-assisted category annotation of the
+KodCode portion, category mapping, and preparation of SFT/RL configurations and
+recorded ablation subsets. TACO and LeetCodeDataset tags originate upstream;
+we organize them into the project's category groups. This release also adds
+a fixed TACO SFT validation split while preserving the historical export.
+
+The underlying problems, reference answers, and tests remain attributed to the
+upstream authors. Source revisions where recovered, frozen file identities,
+and processing notes are recorded under metadata/. TACO's own source
+acknowledgments and the source-specific license terms also apply.

README.md

@@ -1,5 +1,5 @@
 ---
-pretty_name: Code Reasoning Data
+pretty_name: Code Generalization
 language:
 - en
 task_categories:
@@ -8,6 +8,9 @@ tags:
 - code
 - sft
 - reinforcement-learning
+- code-generalization
+- category-annotation
+- ablation
 license: other
 license_name: source-specific-licenses
 license_link: LICENSE
@@ -41,9 +44,34 @@ configs:
     path: data/taco_leetcode/rl/validation.parquet
 ---
 
-# Code Reasoning Data
+# Code Generalization
 
-Code-training data with SFT answers, executable RL tests, and category labels.
+Code Generalization brings together curated data from **KodCode, TACO, and
+LeetCodeDataset** for studying how code training affects generalization across
+tasks and algorithmic categories. We clean and organize the source data, add
+LLM-assisted category annotations to the KodCode portion, and provide SFT and RL
+configurations with reproducible category-ablation subsets.
+
+## What we contribute
+
+- **Data curation.** We select and prepare source problems for training. For
+  KodCode, this includes deduplication by original question, interface repair,
+  static checks, and reference-solution execution checks, yielding 36,074 problems.
+- **Category annotation and organization.** We use an LLM API to classify
+  KodCode problems, retain the classification reasons and review records, and
+  map the reviewed labels into ten categories. For TACO and LeetCodeDataset, we
+  organize the existing upstream tags into seven ablation groups; these are
+  not new API annotations.
+- **Support for controlled experiments.** We provide SFT examples, RL prompts
+  and execution-based reward tests, validation splits, and the recorded sample
+  IDs used for category-removal experiments. This lets users compare training
+  on the full corpus with training on a specified ablation subset.
+
+The underlying problems, reference answers, and tests come from the credited
+upstream datasets. Our work is the curation, additional annotation, category
+mapping, and preparation of these research configurations.
+
+## Data configurations
 
 | Configuration | Train | Validation |
 |---|---:|---:|
@@ -52,29 +80,93 @@ Code-training data with SFT answers, executable RL tests, and category labels.
 | `taco_leetcode_sft` | 21,779 | 100 |
 | `taco_leetcode_rl` | 17,249 | 100 |
 
+Choose an **SFT** configuration for supervised training on reference answers, or
+an **RL** configuration for training with execution-based rewards. KodCode SFT
+and RL share the same problems and split membership; their row counts should
+not be added together as distinct problems. The two data families retain their
+own category systems.
+
+## Quick start
+
+After downloading this repository to a folder named `dataset`:
+
 ```python
 from datasets import load_dataset
 
-ds = load_dataset("./dataset", "kodcode_sft")
-# After upload, replace ./dataset with your Hugging Face dataset repository ID.
+train = load_dataset("./dataset", "kodcode_sft", split="train")
+print(train[0]["messages"])
+
+# Select problems with a particular primary category.
+dp = train.filter(lambda row: row["primary_category"] == "dynamic_programming")
+```
+
+To load from Hugging Face directly, replace `./dataset` with the repository ID
+shown on the dataset page (`owner/Code-generalization`).
+
+| Use | Main fields |
+|---|---|
+| SFT | `messages`, or the text pair `prompt` and `response` |
+| RL | `prompt` (a list of messages) and `reward_model.ground_truth` (tests for the scorer) |
+| Category analysis | `primary_category`; KodCode annotation fields or TACO/LeetCode `overlap_categories` |
+
+Apply the target model's template at training time. Classification reasons and
+reward tests are not SFT targets and should not be appended to training prompts.
+
+## Category annotations
+
+KodCode records include a primary category, auxiliary techniques, a short
+classification reason, and annotation provenance. Across its 36,074 problems,
+7,564 records are marked as reviewed and 1,677 as changed during review.
+The released ten-category labels are a deterministic mapping of the reviewed
+earlier labels. See [KodCode notes](docs/kodcode.md) for the model records and
+field definitions.
+
+A real annotation, showing selected fields:
+
+```json
+{
+  "problem_id": "Algorithm_12700_C",
+  "primary_category": "dynamic_programming",
+  "techniques": ["dp_table", "lcs_reconstruction", "backtracking"],
+  "rationale": "Builds an LCS length table and backtracks to reconstruct the subsequence."
+}
 ```
 
-SFT: use `messages` or `prompt` / `response`. RL: use `prompt` and
-`reward_model.ground_truth`. Apply the target model's template at training time.
-Use only the documented input/target fields; labels and tests are metadata.
+KodCode's main ablations remove the selected **primary category**. TACO/LeetCode
+ablations remove problems when **any mapped category** matches the removed
+group. Use the recorded memberships to reproduce the experimental subsets;
+a simple category filter does not reproduce their additional sampling.
+Export commands are in the [KodCode](docs/kodcode.md) and
+[TACO/LeetCode](docs/taco_leetcode.md) notes.
 
-KodCode uses ten primary categories; TACO/LeetCode uses seven category-removal
-groups. Their taxonomies and historical ablation memberships are separate.
+## Quality and usage notes
 
-The TACO SFT validation split is new in this release. Other splits are historical.
-These configurations are separate experiments, not a joint SFT-to-RL split.
-Historical TACO RL validation overlaps the original SFT pool on 71 problems.
-See [TACO/LeetCode notes](docs/taco_leetcode.md) for historical export.
+- All released KodCode problems carry a historical reference-execution-pass
+  flag. In TACO/LeetCode RL training data, 11,587 records are marked as
+  reference-execution-verified and 5,662 as structural-only. Passing the
+  supplied tests does not establish correctness on every possible input.
+- KodCode's fixed 100-problem validation set comes from Light-RL-10K, whereas
+  its training set combines two KodCode sources. TACO SFT's 100-problem holdout
+  was added for this release; the other splits preserve the historical setup.
+- TACO SFT and RL were separate experiments, not a shared SFT-to-RL partition.
+  There are cross-configuration overlaps; see the
+  [split notes](docs/taco_leetcode.md#splits-and-verification) before combining them.
+- Reward execution requires Linux and an isolated environment. See
+  [verification notes](docs/rl_verification.md) for the two reward rules and
+  execution limits. Loading the dataset does not execute the problem code.
+
+To check the local release, run these commands from the repository root:
 
 ```bash
 pip install -r requirements.txt
 python scripts/validate_release.py --hf
 ```
 
-Source-specific licenses apply; see [LICENSE](LICENSE) and [NOTICE](NOTICE.md).
-Reward scripts require Linux; usage and limits are in [verification notes](docs/rl_verification.md).
+## Sources and license
+
+Please credit the original [KodCode](https://arxiv.org/abs/2503.02951),
+[TACO](https://arxiv.org/abs/2312.14852), and
+[LeetCodeDataset](https://arxiv.org/abs/2504.14655) work when using their data.
+Source repositories and attribution are listed in [NOTICE](NOTICE.md).
+Source-specific terms remain in effect; see [LICENSE](LICENSE), including the
+non-commercial condition on the KodCode-derived portion.

docs/kodcode.md

@@ -1,18 +1,71 @@
 # KodCode
 
-The released pool contains 36,074 distinct problems: 35,974 train and 100 fixed
-validation, shared by SFT and RL. It follows the project's deduplication,
-interface repair, execution checks, and reviewed category mapping.
-The earlier 37,881-answer diagnostic pool is not included as another configuration.
+## Curation
 
-`primary_category` is the ten-category v2 label. `v1_*` fields preserve the
-older reviewed labels; `v1_alternative_category` belongs to that older taxonomy.
-`techniques` retains auxiliary text tags, not extra v2 primary labels.
-Historical main ablations remove the primary category only.
+We prepare a 36,074-problem pool from KodCode-V1-SFT-4o (26,475 problems) and
+KodCode-Light-RL-10K (9,599 problems). The recorded preparation process includes
+deduplication by original question, interface repair, static checks, and
+reference-solution execution checks. All released problems retain a historical
+full-execution-pass flag; this is verification against the supplied tests, not
+a guarantee that the tests cover every valid input.
+
+SFT and RL share 35,974 training and 100 validation problems. The validation
+problems all come from Light-RL-10K; they are not a source-stratified sample of
+the combined pool. The earlier 37,881-answer diagnostic pool is not a separate
+configuration in this release.
+
+## LLM-assisted annotation
+
+We add algorithmic category annotations through an LLM API and retain the
+classification reasons and review history alongside the data. The records
+identify 7,564 reviewed annotations and 1,677 annotations changed during review.
+These counts cover train and validation once, without double-counting the RL
+representation of the same problems.
+
+| Final recorded classifier | Problems |
+|---|---:|
+| `gpt-5.4-mini` | 28,492 |
+| `gpt-5.4` | 7,472 |
+| `gpt-4.1` | 110 |
+
+These are the model identifiers stored in `classifier_model`, not a count of
+all API calls. Review refers to the recorded classification stage; it should
+not be interpreted as independent human verification.
+
+The reviewed v1 primary categories and subcategories are mapped deterministically
+into the ten-category v2 taxonomy. This mapping makes no additional API calls.
+Definitions and mapping rules are in
+[taxonomy.json](../metadata/kodcode/taxonomy.json).
+
+| Fields | Meaning |
+|---|---|
+| `primary_category`, `taxonomy_version` | One final v2 category and its taxonomy version |
+| `v1_primary_category`, `v1_subcategory` | Reviewed labels before the v2 mapping |
+| `v1_alternative_category` | Optional alternative in the older taxonomy, not a second v2 label |
+| `techniques` | Auxiliary free-text techniques, not additional primary labels |
+| `rationale` | The reason for the classification, not the solution's reasoning trace |
+| `classifier_model`, `classification_stage` | Recorded classifier and annotation stage |
+| `was_reviewed`, `review_changed`, `v1_first_pass_category` | Review status and earlier classification |
+| `confidence`, `v1_confidence` | Recorded annotation confidence, not a measured accuracy |
+| `mapping_method` | How the final category was derived |
+
+The original API prompts, request settings, and review-selection logic are not
+included in this release. The annotations can be reused, but the API annotation
+process cannot be rerun end to end from this package alone.
+
+## Reusing the experimental subsets
+
+Historical main ablations remove the primary category only, then select a
+recorded equal-size sample. They do not remove every problem mentioning that
+category in its auxiliary techniques or alternative label.
 
 ```bash
 python scripts/build_kodcode.py --arm no_dynamic_programming --output arm.parquet
 ```
 
-Frozen labeled source exports can be repackaged with `--source DIR`.
-File and membership hashes are in metadata/kodcode/.
+Run the command from the repository root. Use `--kind rl` for the corresponding
+RL subset. The script exports the recorded membership rather than resampling.
+
+`--source DIR` repackages a frozen, already labeled export; it does not rerun the
+upstream cleaning or API annotation pipeline. Source revisions, file hashes,
+and sample memberships are in [metadata/kodcode/](../metadata/kodcode/).

docs/taco_leetcode.md

@@ -1,19 +1,44 @@
 # TACO + LeetCode
 
-The historical SFT pool has 21,879 rows (19,493 TACO; 2,386 LeetCodeDataset).
-This release holds out 100 unique-prompt problems using seed 20260907 and
-source/category stratification, leaving 21,779 training rows.
-The original order and all original SFT fields remain recoverable.
+## Curation and category organization
+
+We bring together filtered TACO and LeetCodeDataset problems, organize their
+upstream tags into seven category-removal groups, and prepare SFT examples and
+RL prompts with reward tests. The category information originates in the
+upstream datasets; this portion was not newly labeled through our LLM API.
+
+`overlap_categories` lists the mapped groups associated with a problem. An
+ablation removes a problem if any of these groups matches the selected category.
+`no_tags` identifies records without mapped tags. The original detailed SFT
+labels remain in `raw_*` columns; the release mapping is in
+[taxonomy.json](../metadata/taco_leetcode/taxonomy.json).
+
+## Splits and verification
+
+The historical SFT pool contains 21,879 rows: 19,493 TACO and 2,386
+LeetCodeDataset. For this release, we hold out 100 unique-prompt problems using
+seed 20260907 and source/category stratification, leaving 21,779 training rows.
+The original SFT fields, membership, and order remain recoverable.
 
 RL retains the original 17,249 train / 100 validation split and all reward tests.
 Its training metadata marks 11,587 reference-execution-verified rows and 5,662
 structural-only rows; validation has 71 and 29 respectively.
-The two task configurations were not built as a shared SFT-to-RL partition.
+These flags describe the recorded checks; structural validation alone does not
+show that a reference solution passes the tests.
+
+There is no train/validation ID or normalized-prompt overlap within each
+configuration, according to the release validation report. However, 71 problems
+in RL validation also occur in SFT training, and 48 problems in the new SFT
+validation set occur in RL training. The configurations preserve separate
+experiments, not a leakage-free shared SFT-to-RL partition.
+
+## Reusing the experimental subsets
 
-Seven removal categories use `overlap_categories`; `no_tags` marks unlabeled
-rows. Original detailed SFT labels remain in `raw_*` columns. Recorded historical SFT
-arms have 14,043 rows and RL arms have 12,168. Release SFT arm exports exclude
-the new holdout and therefore need not remain equal-size.
+The recorded size-matched historical SFT subsets contain 14,043 rows and RL
+subsets contain 12,168. Release SFT subset exports exclude the new holdout and
+therefore need not remain equal-size. Full-corpus configurations have the larger
+counts listed above. Use recorded memberships rather than a new random sample
+when reproducing those experiments.
 
 ```bash
 # Original 21,879-row SFT pool, in original order:
@@ -22,5 +47,8 @@ python scripts/build_taco_leetcode.py --arm full_sft --historical --output origi
 python scripts/build_taco_leetcode.py --kind rl --arm no_graph --output arm.parquet
 ```
 
-The build script also accepts a frozen source export through `--source DIR`.
-Source records, IDs, and split hashes are documented in metadata/taco_leetcode/.
+Run these commands from the repository root. The build script also accepts a
+frozen source export through `--source DIR`; it repackages the prepared data,
+rather than collecting or filtering the upstream datasets from scratch.
+Source records, IDs, and split hashes are in
+[metadata/taco_leetcode/](../metadata/taco_leetcode/).

metadata/release_files.json

@@ -9,10 +9,10 @@
       "bytes": 38,
       "sha256": "6c1257632e97376c4342ba9ce2d0a6a1c84d7c2519a0711a29969970c6e3cd6c"
     },
-    "CHANGELOG.md": {
-      "bytes": 263,
-      "sha256": "145e44726deacf49a583b76c618c287a5659464702929b4bc562dc559669dcf6"
-    },
+    "CHANGELOG.md": {
+      "bytes": 633,
+      "sha256": "08fd8a2a82cbe10cdcbf83b47afce41282eb39a65f399381dc74e9b0315e3443"
+    },
     "data/kodcode/rl/train.parquet": {
       "bytes": 31938297,
       "sha256": "a23ea34d57279ae59d7068c4bbd55df08df49b3d4ca42853a49302365d54f224"
@@ -45,18 +45,18 @@
       "bytes": 195929,
       "sha256": "3f637a2b0f6e466335dddd08834fa56d1eaabf7b1792bf9cbbda21c95501e362"
     },
-    "docs/kodcode.md": {
-      "bytes": 823,
-      "sha256": "f3f28f4e7e6d0d1f0ad3bd435725b893efbc18cf92e27a3554c7ebec5ce7d170"
-    },
+    "docs/kodcode.md": {
+      "bytes": 3465,
+      "sha256": "73cd962b131b9c10cdb8b7db8b97922ed76f15208643fcfcab2d3b0384a7781f"
+    },
     "docs/rl_verification.md": {
       "bytes": 1147,
       "sha256": "958f38e86ff1429bc7063e58d359a39aa2358f1a3c3d2444d42ae7d9539b89c1"
     },
-    "docs/taco_leetcode.md": {
-      "bytes": 1320,
-      "sha256": "d090f3850464245cc2c2d407c23886cba8c9d38573df4c6de85b0c1318dfb3bf"
-    },
+    "docs/taco_leetcode.md": {
+      "bytes": 2729,
+      "sha256": "5e3fa6304d97a9f1ece5738d5b578affaa39d2e7b8e0d7cf455908de2c439a4e"
+    },
     "LICENSE": {
       "bytes": 1094,
       "sha256": "b7b3bf61277c2f6891b3a714ddd0f8dbde92a343f50ac3718afe88c0de2fef18"
@@ -113,14 +113,14 @@
       "bytes": 1802,
       "sha256": "e1aff965903d91f5f489463c4acac910e30be9f5ffc90104e7d3f788f99b5bfd"
     },
-    "NOTICE.md": {
-      "bytes": 823,
-      "sha256": "6480096e84a44838ce4e5d6aaf7dee78be3037bd93a625c032c39cc2fedf4d88"
-    },
-    "README.md": {
-      "bytes": 2397,
-      "sha256": "6ce733786d33625e55f3f4e1c546db0d23f41a3369199c7f52e9445509cabc3d"
-    },
+    "NOTICE.md": {
+      "bytes": 1203,
+      "sha256": "737565466cf2e235747d49e57360e3f0fbca44ec7bbf9a8e106634a091a1033f"
+    },
+    "README.md": {
+      "bytes": 6761,
+      "sha256": "ebd005e1867798e507823cd96e679bf7e92a46d7b2ae0e675af712b07b9753a6"
+    },
     "requirements.txt": {
       "bytes": 46,
       "sha256": "f099c62b8f276a0de63c9b76c820e07b76fa3bb6b4c24c2e1760af1a5e9f4a49"