avewright commited on
Commit
a9a67ed
·
verified ·
1 Parent(s): 7aafc64

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +43 -23
README.md CHANGED
@@ -1,25 +1,45 @@
1
  ---
2
- dataset_info:
3
- features:
4
- - name: messages
5
- list:
6
- - name: role
7
- dtype: string
8
- - name: content
9
- dtype: string
10
- - name: source
11
- dtype: string
12
- - name: bucket
13
- dtype: string
14
- splits:
15
- - name: train
16
- num_bytes: 11383946
17
- num_examples: 6720
18
- download_size: 11208039
19
- dataset_size: 11383946
20
- configs:
21
- - config_name: default
22
- data_files:
23
- - split: train
24
- path: data/train-*
25
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ license: apache-2.0
3
+ task_categories:
4
+ - text-generation
5
+ language:
6
+ - en
7
+ tags:
8
+ - chat
9
+ - sft
10
+ - math
11
+ - coding
12
+ - agents
13
+ - instruction-following
14
+ size_categories:
15
+ - 1K<n<10K
 
 
 
 
 
 
 
 
 
16
  ---
17
+
18
+ # downsize-macd
19
+
20
+ Curated multi-turn / single-turn chat SFT mix for **downsize** distillation:
21
+
22
+ | bucket | intent |
23
+ |---|---|
24
+ | `math` | word problems + concise solutions |
25
+ | `coding` | code instruct (SmolTalk self-oss) |
26
+ | `agentic` | Fable / AgentTrove / API-gen traces |
27
+ | `direction` | system following + short helpful seeds |
28
+
29
+ ## Stream
30
+
31
+ ```python
32
+ from datasets import load_dataset
33
+ ds = load_dataset("avewright/downsize-macd", split="train", streaming=True)
34
+ for row in ds:
35
+ print(row["bucket"], row["messages"][0]["content"][:80])
36
+ break
37
+ ```
38
+
39
+ ## Columns
40
+
41
+ - `messages`: list of `{role, content}`
42
+ - `source`: upstream dataset id
43
+ - `bucket`: math | coding | agentic | direction
44
+
45
+ Built with concise-truth filters (drop loops / unclosed think / overlong waffle).