Publish cleaned verl parquet (39179 rows) + cleaning report

README.md

@@ -8,61 +8,64 @@ tags:
   - verl
   - reinforcement-learning
 size_categories:
-  - 1K<n<10K
+  - 10K<n<100K
 configs:
   - config_name: default
     data_files:
       - split: train
-        path: data/train.jsonl
-      - split: validation
-        path: data/val.jsonl
+        path: data/verl_deepscaler_clean_39179.parquet
 ---
 
 # DeepScaleR-Verl-Clean
 
-Cleaned, deduplicated export of the **agentica-org/DeepScaleR-Preview-Dataset**
-(https://huggingface.co/datasets/agentica-org/DeepScaleR-Preview-Dataset), formatted for
-verl-based GRPO training. Source dataset is MIT-licensed; this derived subset keeps the
-same license.
+A single cleaned, deduplicated, verl-ready parquet built from four partial dumps of the
+DeepScaleR math dataset (`agentica-org/DeepScaleR-Preview-Dataset`, MIT licensed). It is
+meant for rule-based-reward GRPO/RL training with the
+[verl](https://github.com/volcengine/verl) framework.
 
-## Splits
+## Files
 
-| Split | File | Rows |
-|---|---|---|
-| Train | `data/train.jsonl` | 1950 |
-| Validation | `data/val.jsonl` | 50 |
+| File | Rows |
+|---|---|
+| `data/verl_deepscaler_clean_39179.parquet` | 39179 |
+| `data/cleaning_report.json` | full paper trail of every dropped row |
 
-## Cleaning
+## How it was built
 
-The raw export contained 2100 lines. Lines were rejected in the following order
-(first match wins), leaving 2000 usable rows:
+Sources were read in priority order (`deepscaler_part1.jsonl`, `deepscaler_part2.jsonl`,
+`deepscaler_export.json`, `deepscaler_legacy.jsonl`; within a file, earlier rows win):
 
-1. **`foreign_source`** — 15 lines whose `data_source` was not `DeepScaleR`
-   (entries that slipped in from the merged pool).
-2. **`missing_field`** — 25 lines whose `answer` or `solution` was absent or blank.
-3. **`duplicate`** — 60 lines whose problem text already appeared on an earlier
-   surviving line; only the first occurrence is kept.
+1. **Validity** - a row is garbage unless its statement *and* its answer are both present
+   and non-blank after trimming. Invalid rows are removed *before* deduplication, so a
+   broken copy can never steal the slot of a good copy. Rows whose statement is unusable
+   are recorded as `missing_problem`, rows with a usable statement but no answer as
+   `missing_answer`. A missing `solution` is not garbage: it is stored as an empty string.
+2. **Identity / deduplication** - identity is the trimmed statement text. Only the first
+   occurrence (by source priority, then row order) is kept; every later copy is dropped as
+   `duplicate`, including legacy copies that carry stale answers.
+3. **Ordering / indexing** - survivors are sorted by statement text ascending (plain
+   code-point ordering) and numbered 0..N-1; `extra_info.index` equals the row's position
+   and the parquet is physically in that order. Prompts carry the trimmed statement and
+   `reward_model.ground_truth` the trimmed answer.
 
-Surviving rows were sorted by problem text (ascending, plain string comparison, ignoring
-surrounding whitespace), numbered 0..1999, and every row whose index is divisible by 40
-was assigned to validation, the rest to train. Row order inside each file is ascending
-`extra_info.index`. Problems, answers and solutions are carried over verbatim from the
-source - no reformatting or re-derivation.
+Totals: 43608 input rows -> 39179 kept, 4390 dropped as duplicates, 39 dropped as invalid.
+See `data/cleaning_report.json` for the per-row detail of everything that was thrown away.
 
 ## Schema (verl)
 
-Each JSON Lines row:
+Five columns, in this order:
 
-```json
-{
-  "data_source": "DeepScaleR",
-  "prompt": [{"role": "user", "content": "<problem text>"}],
-  "ability": "math",
-  "reward_model": {"style": "rule", "ground_truth": "<answer>"},
-  "extra_info": {"index": 0, "solution": "<reference solution>"}
-}
-```
+| Column | Content |
+|---|---|
+| `data_source` | `"DeepScaleR"` |
+| `prompt` | `[{"role": "user", "content": "<trimmed problem statement>"}]` |
+| `ability` | `"math"` |
+| `reward_model` | `{"style": "rule", "ground_truth": "<trimmed answer>"}` |
+| `extra_info` | `{"index": <row position>, "solution": "<solution, or empty string>"}` |
+
+## Usage
 
-See `dataset_report.json` in this repository for the full accounting of the cleaning run,
-including the rejection counts per reason and the SHA-256 digests of the parquet files that
-back this dataset.
+```python
+from datasets import load_dataset
+ds = load_dataset("dusersad12/DeepScaleR-Verl-Clean", split="train")
+```

data/verl_deepscaler_clean_39179.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b5c72a5e8b4dcd5203845a632d56f0a8f306b87c2f6cef3d6cafd453ebe71748
+size 6523404

dataset_report.json

@@ -1,29 +0,0 @@
-{
-  "raw_rows": 2100,
-  "kept_rows": 2000,
-  "dropped": {
-    "foreign_source": 15,
-    "missing_field": 25,
-    "duplicate": 60
-  },
-  "train_rows": 1950,
-  "val_rows": 50,
-  "manifest_check": {
-    "discrepancies": [
-      {
-        "field": "valid_unique_problems",
-        "manifest_value": 2050,
-        "observed_value": 2000
-      },
-      {
-        "field": "known_duplicate_entries",
-        "manifest_value": 75,
-        "observed_value": 60
-      }
-    ]
-  },
-  "sha256": {
-    "verl_deepscaler_train.parquet": "e51b4f3dc20da2a8eef9c75797a2df481510e4232983041daa74dc551bd529fa",
-    "verl_deepscaler_val.parquet": "83e0271fe9c292a4de453f11ea7936f081796503ed0986649cea3e46e65cc348"
-  }
-}