Datasets:
batch 1: upload SCHEMA.md (19 execution-verified Python examples)
Browse files
SCHEMA.md
ADDED
|
@@ -0,0 +1,59 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Unified Training Example Schema (v1)
|
| 2 |
+
|
| 3 |
+
Every example that enters the dataset follows EXACTLY this schema (JSON / parquet row):
|
| 4 |
+
|
| 5 |
+
```json
|
| 6 |
+
{
|
| 7 |
+
"id": "taco-0000123", // unique: <source-short>-<row-id>
|
| 8 |
+
"source_dataset": "BAAI/TACO", // provenance
|
| 9 |
+
"source_url": "https://huggingface.co/datasets/BAAI/TACO",
|
| 10 |
+
"source_row_id": "123", // id/idx in the source row
|
| 11 |
+
"language": "python", // ALWAYS python (hard rule)
|
| 12 |
+
"task_type": "stdin_program | function | tir_math | swe",
|
| 13 |
+
|
| 14 |
+
"instruction": "...", // what a REAL user would have said (see RULES.md §2)
|
| 15 |
+
"chain_of_thought": "...", // reasoning trace BEFORE the code (see RULES.md §1)
|
| 16 |
+
"code": "...", // final verified Python solution
|
| 17 |
+
"doc": "...", // short clear explanation of what the code does
|
| 18 |
+
|
| 19 |
+
"tests": { // the executable proof
|
| 20 |
+
"kind": "stdin_stdout | pytest | tir_answer_check",
|
| 21 |
+
"cases": [ {"input": "...", "expected_output": "..."} ],
|
| 22 |
+
"num_cases": 5
|
| 23 |
+
},
|
| 24 |
+
|
| 25 |
+
"messages": [ // ready-to-train chat rendering
|
| 26 |
+
{"role": "user", "content": "<instruction>"},
|
| 27 |
+
{"role": "assistant", "content": "<chain_of_thought wrapped in <think>> + code block + doc"}
|
| 28 |
+
],
|
| 29 |
+
|
| 30 |
+
"token_counts": { // tiktoken o200k_base, counted programmatically
|
| 31 |
+
"instruction": 0, "chain_of_thought": 0, "code": 0, "doc": 0,
|
| 32 |
+
"tests": 0, "total_training_fields": 0, // instruction+cot+code+doc
|
| 33 |
+
"total_with_tests": 0
|
| 34 |
+
},
|
| 35 |
+
"context_bucket": "under_16k | over_16k", // by total_training_fields
|
| 36 |
+
|
| 37 |
+
"verification": {
|
| 38 |
+
"executed": true,
|
| 39 |
+
"all_tests_passed": true,
|
| 40 |
+
"passed_on_attempt": 1, // repairs attempted (max 2, then discard)
|
| 41 |
+
"code_modified_from_source": false, // did we fix/improve the source code?
|
| 42 |
+
"cot_source": "generated | native_verified | native_rewritten",
|
| 43 |
+
"cot_quality_checks": ["faithful_to_code","has_edge_case_analysis","has_self_correction","ends_with_verification"],
|
| 44 |
+
"worth_adding_reason": "one-line justification"
|
| 45 |
+
},
|
| 46 |
+
"created_at": "2026-10-03"
|
| 47 |
+
}
|
| 48 |
+
```
|
| 49 |
+
|
| 50 |
+
## Folder layout
|
| 51 |
+
```
|
| 52 |
+
data/under_16k/ -> verified examples <= 16,384 tokens (THE dataset; uploaded to HF)
|
| 53 |
+
data/over_16k/ -> verified examples > 16,384 tokens (kept separately, not uploaded yet)
|
| 54 |
+
data/discarded/ -> rejected candidates + reason log (audit trail)
|
| 55 |
+
work/ -> intermediate candidate pools per source dataset
|
| 56 |
+
```
|
| 57 |
+
|
| 58 |
+
## HF upload target
|
| 59 |
+
`dsfasdfaserwer/Dataset` — files: `README.md` (card), `data/train-00000-of-00001.parquet`, `SCHEMA.md`, `RULES.md`, `VALIDATION_LOG.md`
|