bunnycore commited on
Commit
6ac5f21
·
verified ·
1 Parent(s): e41d926

Add dataset card (format, composition, verification, limitations)

Browse files
Files changed (1) hide show
  1. README.md +111 -0
README.md ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-4.0
3
+ language:
4
+ - en
5
+ task_categories:
6
+ - text-generation
7
+ tags:
8
+ - code
9
+ - distillation
10
+ - synthetic
11
+ - reasoning
12
+ - sft
13
+ size_categories:
14
+ - n<1K
15
+ configs:
16
+ - config_name: default
17
+ data_files:
18
+ - split: train
19
+ path: distillation-150.parquet
20
+ ---
21
+
22
+ # distillation-150
23
+
24
+ 150 single-turn (user → assistant) reasoning-focused coding samples for supervised
25
+ fine-tuning (SFT) / distillation of small language models. Every assistant reply
26
+ starts with a structured `<think>` block that derives the solution from first
27
+ principles instead of pattern-completing a memorized template, followed by a concise
28
+ explanation and a complete, runnable program.
29
+
30
+ ## Format
31
+
32
+ Single Parquet file (`distillation-150.parquet`, zstd compression), one row per sample,
33
+ schema:
34
+
35
+ ```
36
+ messages: list<struct<role: string, content: string>>
37
+ ```
38
+
39
+ Each `messages` is `[user, assistant]`. The user turn is a coding task/brief; the
40
+ assistant turn is:
41
+
42
+ ```
43
+ <think>
44
+ 1. DECONSTRUCTION & INVARIANTS: ...
45
+ 2. ARCHITECTURE & BINDING CONTRACT: ... (INTERFACE & STRUCTURE for Python)
46
+ 3. DRAFTING CORE LOGIC & SELF-CORRECTION: ...
47
+ 4. CONCRETE DRY-RUN TRACE: ...
48
+ 5. COMPLETENESS VERIFICATION: ...
49
+ </think>
50
+
51
+ <2–4 sentence explanation>
52
+
53
+ <complete runnable code in one fenced block>
54
+ ```
55
+
56
+ ## Composition
57
+
58
+ | Slice | Count | Language | Delivery |
59
+ |---|---|---|---|
60
+ | Games | 25 | HTML/JS (single file, vanilla) | canvas games: breakout, lane-dodge, 2048-lite, minesweeper, … |
61
+ | Apps | 50 | HTML/JS (single file, vanilla) | timers, calculators, converters, localStorage tools |
62
+ | Interactive tools/viz | 25 | HTML/JS + Canvas + Web Audio | sort/pathfinding visualizers, Web Audio synths |
63
+ | Algorithms | 50 | Python 3 (stdlib only) | DP, parsing, graphs, heaps, with `if __name__ == "__main__"` demos |
64
+
65
+ No external libraries, fonts, images, CDNs, or assets anywhere in the samples.
66
+ `localStorage` access is wrapped in `try/catch`; `preventDefault` on game keys.
67
+
68
+ ## Loading
69
+
70
+ ```python
71
+ from datasets import load_dataset
72
+
73
+ ds = load_dataset("bunnycore/distillation-150", split="train")
74
+ print(ds) # Dataset({features: ['messages'], num_rows: 150})
75
+ print(ds[0]["messages"][0]["content"][:200])
76
+ ```
77
+
78
+ Or directly:
79
+
80
+ ```python
81
+ import pandas as pd
82
+ df = pd.read_parquet("distillation-150.parquet")
83
+ ```
84
+
85
+ ## Verification performed
86
+
87
+ Every row was machine-checked before upload:
88
+
89
+ - `<think>…</think>` tags present; exactly one fenced code block per assistant turn;
90
+ no `TODO`/`FIXME` markers in code.
91
+ - HTML entries: all `<script>` blocks pass `node --check`; every
92
+ `getElementById('x')` resolves to an `id="x"` in the same entry's markup.
93
+ - Python entries: code blocks pass `ast.parse`; every embedded
94
+ `if __name__ == "__main__":` demo **executes** and prints the annotated expected output.
95
+ - No duplicate topics or duplicate user prompts; think blocks ≈ 280–630 words.
96
+
97
+ ## Honest limitations
98
+
99
+ - **Not human-reviewed.** These are machine-generated and machine-checked samples.
100
+ No browser/e2e testing was run on the 100 HTML/JS entries; canvas/audio behavior is
101
+ traced, not observed.
102
+ - Two near-duplicate mechanics (`stroop-test`, `color-reaction`) slipped through review.
103
+ - Think-block length deviates from the ~350–550 word target in both directions.
104
+ - Small sample (n=150) — a seed set to iterate on, not a training-ready corpus.
105
+ Expect several percent of entries to need manual repair before training.
106
+ - Assistant code follows the stated constraints (single file, stdlib-only) and encodes
107
+ one reference style; style diversity is low by construction.
108
+
109
+ ## License
110
+
111
+ CC-BY-4.0.