Upload README.md with huggingface_hub

README.md

@@ -1,51 +1,55 @@
 ---
-dataset_info:
-  features:
-  - name: task_id
-    dtype: string
-  - name: difficulty
-    dtype: int64
-  - name: difficulty_tier
-    dtype: string
-  - name: n_turns
-    dtype: int64
-  - name: source_agent
-    dtype: string
-  - name: messages
-    list: json
-  - name: tools
-    list:
-    - name: type
-      dtype: string
-    - name: function
-      struct:
-      - name: name
-        dtype: string
-      - name: description
-        dtype: string
-      - name: parameters
-        struct:
-        - name: type
-          dtype: string
-        - name: properties
-          struct:
-          - name: command
-            struct:
-            - name: type
-              dtype: string
-            - name: description
-              dtype: string
-        - name: required
-          list: string
-  splits:
-  - name: train
-    num_bytes: 29097476
-    num_examples: 4677
-  download_size: 28918861
-  dataset_size: 29097476
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
+license: mit
+task_categories:
+- text-generation
+tags:
+- sft
+- agent
+- tool-calling
+- data-analysis
+- trl
 ---
+
+# 🛠️ Data Agent — SFT
+
+**4,677 worked examples** of an agent doing data science *the right way*. Each row is a complete,
+**verified-correct** trajectory: read the question, poke at the data with a shell tool, reason,
+compute, and write the answer. Every one of these solved its task and passed a deterministic grader
+— so you're fine-tuning on demonstrations that are **known to be correct**, not just plausible.
+
+Drop-in ready for [TRL](https://github.com/huggingface/trl): conversational `messages` + `tools`.
+
+## Where it comes from
+These are real agent rollouts on the [Data Agent](https://huggingface.co/HuggingEnvs) tasks, which
+were themselves built from the [**jupyter-agent dataset**](https://huggingface.co/datasets/jupyter-agent/jupyter-agent-dataset)
+(data-science notebooks over Kaggle datasets). We kept **only trajectories that reached the correct
+answer** under deterministic grading (reward = 1.0) — one clean demonstration per task.
+
+## What's inside
+- **4,677 correct trajectories** — one per task
+- **Difficulty** — easy 1,402 · medium 2,640 · hard 635
+- One tool throughout: `bash` (shell command execution)
+
+## What's in a row
+- **`messages`** — the full conversation in OpenAI/TRL chat format: `system` → `user` (the task) →
+  `assistant` (reasoning + `tool_calls`) → `tool` (command output) → … → final `assistant` answer.
+  Tool-call `arguments` are JSON objects; `tool` messages carry the tool `name`.
+- **`tools`** — the `bash` tool's JSON schema (rendered by `apply_chat_template(..., tools=...)`)
+- `task_id`, `difficulty` (1–5), `difficulty_tier`, `n_turns`, `source_agent`
+
+## Fine-tune with TRL
+```python
+from datasets import load_dataset
+from trl import SFTTrainer, SFTConfig
+
+ds = load_dataset("HuggingEnvs/data-agent-sft", split="train")
+
+trainer = SFTTrainer(
+    model="Qwen/Qwen2.5-3B-Instruct",     # any tool-capable chat template
+    train_dataset=ds,                      # messages + tools are picked up automatically
+    args=SFTConfig(assistant_only_loss=True, max_length=8192),
+)
+trainer.train()
+```
+The `messages` + `tools` columns render through your model's chat template, and
+`assistant_only_loss=True` trains on the assistant's tokens only — no dataset wrangling needed.