Upload README.md with huggingface_hub

README.md

@@ -1,68 +1,27 @@
 ---
 license: apache-2.0
-task_categories:
-- text-generation
-language:
-- en
-tags:
-- chat
-- sft
-- math
-- coding
-- agents
-- instruction-following
-size_categories:
-- 1K<n<10K
-dataset_info:
-  features:
-  - name: messages
-    list:
-    - name: role
-      dtype: string
-    - name: content
-      dtype: string
-  - name: source
-    dtype: string
-  - name: bucket
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 16332509
-    num_examples: 8200
-  download_size: 15972059
-  dataset_size: 16332509
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
+task_categories: [text-generation]
+language: [en]
+tags: [chat, sft, math, coding, agents, instruction-following, distillation]
+size_categories: [1K<n<10K]
 ---
 
 # downsize-macd
 
-Curated multi-turn / single-turn chat SFT mix for **downsize** distillation:
+Multi-source curated chat SFT for small-student distillation (**math / coding / agentic / direction**).
 
-| bucket | intent |
-|---|---|
-| `math` | word problems + concise solutions |
-| `coding` | code instruct (SmolTalk self-oss) |
-| `agentic` | Fable / AgentTrove / API-gen traces |
-| `direction` | system following + short helpful seeds |
+Built to favor **concise correct answers** over long fluent-wrong CoT. 20+ upstream sources.
 
 ## Stream
 
 ```python
 from datasets import load_dataset
 ds = load_dataset("avewright/downsize-macd", split="train", streaming=True)
-for row in ds:
-    print(row["bucket"], row["messages"][0]["content"][:80])
-    break
+row = next(iter(ds))
+print(row["bucket"], row["source"], row["messages"][0]["content"][:120])
 ```
 
 ## Columns
-
-- `messages`: list of `{role, content}`
-- `source`: upstream dataset id
+- `messages`: chat turns `{role, content}`
+- `source`: upstream dataset
 - `bucket`: math | coding | agentic | direction
-
-Built with concise-truth filters (drop loops / unclosed think / overlong waffle).