dsfasdfaserwer commited on
Commit
3159f3d
·
verified ·
1 Parent(s): 1c738c8

batch 1: upload README.md (19 execution-verified Python examples)

Browse files
Files changed (1) hide show
  1. README.md +67 -0
README.md CHANGED
@@ -1,3 +1,70 @@
1
  ---
2
  license: apache-2.0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: apache-2.0
3
+ task_categories:
4
+ - text-generation
5
+ - question-answering
6
+ tags:
7
+ - code
8
+ - python
9
+ - reasoning
10
+ - chain-of-thought
11
+ - execution-verified
12
+ - tir
13
+ pretty_name: Python Verified 16k (batch 1)
14
+ size_categories:
15
+ - n<1K
16
+ configs:
17
+ - config_name: default
18
+ data_files:
19
+ - split: train
20
+ path: data/train-*.parquet
21
  ---
22
+
23
+ # Python-Verified-16k — execution-verified Python training examples (batch 1)
24
+
25
+ **19 examples · 100% Python · 100% execution-verified · all ≤ 16,384 tokens (o200k_base)**
26
+
27
+ Every example follows one unified schema (see `SCHEMA.md`) and passed a hard verification gate
28
+ (see `RULES.md` §4): the code **actually ran** and passed its tests in a sandboxed subprocess
29
+ (timeout + memory + nproc limits), the chain-of-thought is faithful to the code, the instruction
30
+ is a realistic user request that does **not** leak implementation details, and a short doc
31
+ explains the solution.
32
+
33
+ ## Batch 1 composition
34
+
35
+ | Source | Examples | Verification performed |
36
+ |---|---|---|
37
+ | BAAI/TACO (competitive programming) | 11 | solutions **rewritten for clarity**, then re-run against the **FULL original test suites** (1,794+ test executions in total, e.g. 348/348, 278/278, 202/202 per problem) |
38
+ | AI-MO/NuminaMath-TIR (tool-integrated math reasoning) | 8 | native GPT-4 TIR traces **re-executed**: every embedded Python block ran and the printed output matched the boxed answer |
39
+
40
+ Token counts (instruction+CoT+code+doc): min 548, max 1778, mean 931.
41
+ Bucketing: examples ≤ 16,384 tokens live in `data/` (this repo); > 16,384-token examples are kept
42
+ in a separate local `over_16k/` folder (none in batch 1).
43
+
44
+ ## Fields
45
+
46
+ `id`, `source_dataset`, `source_url`, `source_row_id`, `language` (always `python`), `task_type`
47
+ (`stdin_program` | `tir_math`), **`instruction`** (natural user request), **`chain_of_thought`**
48
+ (reasoning before/around the code; TIR examples interleave reasoning and Python), **`code`**
49
+ (verified solution), **`doc`** (what the code does), `tests` (JSON: executed test cases + counts),
50
+ `messages` (JSON: ready-to-train chat rendering), `token_counts` (JSON), `context_bucket`,
51
+ `verification` (JSON: full audit trail), `created_at`.
52
+
53
+ ## Provenance & licenses
54
+
55
+ - Derived from [BAAI/TACO](https://huggingface.co/datasets/BAAI/TACO) (Apache-2.0) and
56
+ [AI-MO/NuminaMath-TIR](https://huggingface.co/datasets/AI-MO/NuminaMath-TIR) (Apache-2.0).
57
+ Both sources are Apache-2.0, so this derived dataset is released as Apache-2.0.
58
+ Authored instructions, chains-of-thought, docs and code rewrites in this repo are original work.
59
+ - Full audit trail: `VALIDATION_LOG.md` · Schema: `SCHEMA.md` · Quality rules: `RULES.md`.
60
+
61
+ ## How it was made (pipeline)
62
+
63
+ 1. Source rows streamed from the top-tier training datasets (quality-ranked list, training-only, Python-mandatory).
64
+ 2. Bundled solutions executed against bundled tests in a sandbox (subprocess, 8s CPU / 10s wall,
65
+ 1.6 GB RSS cap, no env inheritance). Failures: bounded repair (max 2 attempts) else discard — no infinite loops.
66
+ 3. Instructions authored as realistic user requests (full functional spec, **no** prescribed
67
+ libraries/algorithms/APIs). CoTs follow researched best practice: restate → edge cases →
68
+ alternatives → commit + justify → derive code → self-correction → concrete dry-run → complexity.
69
+ 4. Every rewritten/renormalized artifact **re-verified against the full original test suite** before acceptance.
70
+ 5. Tokenized with tiktoken `o200k_base`; ≤16,384 → this dataset, >16,384 → separate bucket.