Datasets:

ecnu-icalk
/

SkillGym

@@ -1,203 +1,294 @@
 ---
 pretty_name: SkillGym
 license: other
-license_name: skillgym-mixed-licensing
-license_link: https://github.com/ECNU-ICALK/SkillGym/blob/main/LICENSE
 language:
-  - en
 source_datasets:
-  - original
 tags:
-  - ai-agents
-  - benchmark
-  - skills
-  - task-generation
-  - agent-evaluation
-  - text
 ---
 <div align="center">
-  <h1>SkillGym Dataset</h1>
-  <p><strong>Executable tasks, human-written skills, and verified long-horizon agent trajectories.</strong></p>
-  <p><a href="https://github.com/ECNU-ICALK/SkillGym">GitHub / code</a> · <a href="https://huggingface.co/ecnu-icalk/SkillGym-Agent">Agent checkpoint</a> · <a href="https://github.com/ECNU-ICALK/SkillGym/blob/main/assets/Internalizing_Large_Scale_Human_Written_Skills_into_LLMs_for_Real_World_Problem_Solving.pdf">Paper</a></p>
 </div>
-SkillGym is a skill-grounded benchmark and training resource. It instantiates human-written skills and reusable task templates as executable environments, checks outcomes with code-based verifiers, and measures skill dependence through paired executions with and without the target skill.
-## At a glance
-| Artifact | Contents | Release size |
-| --- | --- | ---: |
-| `Tasks.tar.zst` | Paired executable task environments | 5,512 variants / 9 GB |
-| `skill_library.tar.zst` | Human-written skills and supporting assets | 117 MB |
-| `task_templates.tar.zst` | Reusable seed templates and fixtures | 63 templates / 110 MB |
-| `Trajectories/*.jsonl` | Successful trajectories grouped by result type, harness, and teacher | 8,364 records / 8 files |
-The release contains **2,756 accepted environments** across **12 major categories** and **63 sub-categories**. The archive sizes are approximate and can change with Hub storage metadata.
-## Quick start
-Install the Hub CLI and download only the artifacts you need:
 ```bash
 python -m pip install -U huggingface_hub datasets
-# Skills and templates for task construction
 hf download ecnu-icalk/SkillGym \
-  skill_library.tar.zst task_templates.tar.zst \
-  --repo-type dataset --local-dir .hf/skillgym
-# Optional: published environments (the largest archive)
-hf download ecnu-icalk/SkillGym Tasks.tar.zst \
-  --repo-type dataset --local-dir .hf/skillgym
-# Optional: all trajectory collections
-hf download ecnu-icalk/SkillGym \
-  --include "Trajectories/*.jsonl" \
-  --repo-type dataset --local-dir .hf/skillgym
-tar --zstd -xf .hf/skillgym/skill_library.tar.zst
-tar --zstd -xf .hf/skillgym/task_templates.tar.zst
-tar --zstd -xf .hf/skillgym/Tasks.tar.zst
 ```
-To download one trajectory file instead of the full collection:
 ```bash
 hf download ecnu-icalk/SkillGym \
-  Trajectories/skill_dependent_claude_code_deepseek_v4_pro.jsonl \
-  --repo-type dataset --local-dir .hf/skillgym
 ```
-## Contents and task definition
-### Directory layout
 ```text
-Tasks/
-├── Skill-Dep/
-│   ├── withskill/
-│   └── withoutskill/
-└── Verifier-Passed/
-    ├── withskill/
-    └── withoutskill/
-skill_library/
-  <skill>/...
-task_templates/
-  <major>/<minor>/seed_task/...
-Trajectories/
-  skill_dependent_*.jsonl
-  verifier_passed_fallback_*.jsonl
 ```
-A task variant is organized as:
-```text
-Tasks/<result>/<skill_condition>/<major>/<minor>/<skill>/<taskN>/
-├── task.toml
-├── instruction.md
-├── environment/
-├── tests/
-└── solution/
-```
-In prose, we call these the **with-skill** and **without-skill** conditions. The archive directories are named `withskill` and `withoutskill`: the former installs the target skill, while the latter keeps the task assets but removes it. A pair is labeled **Skill-Dep.** when the reference with-skill execution succeeds and the paired without-skill execution produces a valid reward failure. **Verifier-Passed** contains valid verifier-backed tasks that do not meet that stricter contrastive criterion. These are construction-time labels, not universal claims about skill necessity.
-### Task and trajectory statistics
-| Measure | Value |
-| --- | ---: |
-| Accepted environments | 2,756 |
-| Skill-Dep. / Verifier-Passed | 1,081 / 1,675 |
-| Paired variants | 5,512 |
-| Successful trajectories | 8,364 from 48,152 trials |
-| Trial-level success rate | 17.4% |
-| Unique tasks after deduplication | 2,302 |
-| Average successful trajectory | 49.0 tool calls · 63.4k logged text tokens · 35.2 interaction steps |
-| Longest observed trajectory | 350 tool calls · 342.9k logged text tokens · 318 interaction steps |
-The logged text-token count describes stored trajectory text. It is not API usage, latent reasoning, or the context length of one model request.
-## Load trajectory records
-Each JSONL line is one trajectory record. Streaming avoids materializing the full corpus:
-```python
-from datasets import load_dataset
-records = load_dataset(
-    "json",
-    data_files={
-        "train": "hf://datasets/ecnu-icalk/SkillGym/Trajectories/skill_dependent_claude_code_deepseek_v4_pro.jsonl",
-    },
-    split="train",
-    streaming=True,
-)
-for record in records.take(1):
-    print(record["session_id"])
-```
-Most files use this normalized schema:
-```json
-{
-  "session_id": "...",
-  "path_index": 0,
-  "num_messages": 28,
-  "model": "...",
-  "messages": [],
-  "tools": [],
-  "provenance": {}
-}
-```
-The six normalized files expose `session_id`, `path_index`, `num_messages`, `model`, `messages`, `tools`, and `provenance`. The two `*_codex_nex_n2_pro.jsonl` files preserve their source Codex trace schema, so top-level fields are not identical across all eight files. Inspect the selected file before concatenating records.
-Trajectory filenames follow:
-```text
-<result_type>_<harness>_<teacher_model>.jsonl
-```
-- `skill_dependent_...jsonl`: trajectories associated with Skill-Dep. tasks.
-- `verifier_passed_fallback_...jsonl`: trajectories associated with Verifier-Passed fallback tasks.
-- The middle component identifies the harness, such as `claude_code` or `codex`.
-- The final component identifies the teacher-model label.
-The records are model-generated execution logs. Interpret them together with the task environment and verifier rather than as independent human demonstrations. Because the two Codex-Nex files contain nested fields that differ from the normalized files, automatic dataset viewers may fail when all eight files are combined; load each file separately or normalize the records before concatenation.
-The eight release files are grouped as follows:
-| Result type | Harness | Teacher labels | Files |
-| --- | --- | --- | ---: |
-| `skill_dependent` | Claude Code | DeepSeek V4 Pro, GLM-5.2 | 2 |
-| `skill_dependent` | Codex | GPT-5.4, Nex-N2-Pro | 2 |
-| `verifier_passed_fallback` | Claude Code | DeepSeek V4 Pro, GLM-5.2 | 2 |
-| `verifier_passed_fallback` | Codex | GPT-5.4, Nex-N2-Pro | 2 |
-## Dataset provenance
-The archives are the large-artifact snapshot associated with GitHub commit `6ebabba`. The [GitHub repository](https://github.com/ECNU-ICALK/SkillGym) remains the source for the task builder, documentation, paper materials, and lightweight figures. The [SkillGym-Agent model card](https://huggingface.co/ecnu-icalk/SkillGym-Agent) documents the checkpoint trained on these trajectories.
-## Licensing and responsible use
-The GitHub repository and repository-level materials are released under the [MIT License](https://github.com/ECNU-ICALK/SkillGym/blob/main/LICENSE). This dataset uses mixed licensing: source skills, fixtures, and supporting assets may retain upstream notices or additional terms. Preserve the notices included with extracted files before redistribution.
-The release is intended for research on long-horizon agents, tool-mediated workflows, procedural knowledge, and verified trajectory training. Review executable environments and instructions before running them in a privileged or networked environment. A Skill-Dep. label is empirical evidence under the reference construction setup; it is not a safety guarantee or a measure of universal skill necessity.
 ## Citation
-The accompanying paper is under double-blind review. Until final publication metadata is available, use:
 ```bibtex
-@misc{skillgym,
-  title = {SkillGym: Internalizing Human Skills into LLMs for Real-World Problem Solving},
   year  = {2026},
-  note  = {Under review at ICLR 2027},
   url   = {https://github.com/ECNU-ICALK/SkillGym}
 }
 ```

 ---
 pretty_name: SkillGym
 license: other
 language:
+- en
 source_datasets:
+- original
 tags:
+- ai-agents
+- benchmark
+- skills
+- task-generation
+- agent-evaluation
+- long-horizon
+- tool-use
+- verifier
+- text
 ---
 <div align="center">
+<picture>
+  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/ECNU-ICALK/SkillGym/main/assets/brand-dark.png">
+  <source media="(prefers-color-scheme: light)" srcset="https://raw.githubusercontent.com/ECNU-ICALK/SkillGym/main/assets/brand-light.png">
+  <img src="https://raw.githubusercontent.com/ECNU-ICALK/SkillGym/main/assets/brand-light.png" alt="SkillGym" width="520">
+</picture>
+### SkillGym Dataset
+**Human-written skills, executable environments, and verified long-horizon agent trajectories**
+[GitHub / Code](https://github.com/ECNU-ICALK/SkillGym) ·
+[SkillGym-Agent](https://huggingface.co/ecnu-icalk/SkillGym-Agent) ·
+[Paper](https://github.com/ECNU-ICALK/SkillGym/blob/main/assets/Internalizing_Large_Scale_Human_Written_Skills_into_LLMs_for_Real_World_Problem_Solving.pdf)
 </div>
+## Dataset Summary
+**SkillGym** is the data release accompanying the SkillGym framework. It connects human-written
+procedural knowledge to executable task environments, code-based outcome verifiers, paired
+with-skill / without-skill executions, and long-horizon agent trajectories.
+The dataset is intended to support research on skill-grounded task construction, agent training,
+verifier-based evaluation, procedural transfer, and outcome-based learning.
+| Statistic | Value |
+| --- | ---: |
+| **Accepted environments** | **2,756** |
+| **Paired task variants** | **5,512** |
+| **Skill-Dep. / Verifier-Passed** | **1,081 / 1,675** |
+| **Successful trajectories** | **8,364** |
+| **Sampled trials** | **48,152** |
+| **Overall sampled-trial success rate** | **17.4%** |
+| **Deduplicated task coverage** | **2,302 unique tasks** |
+| **Taxonomy** | **12 major / 63 sub-categories** |
+Successful trajectories cover 62 of the 63 sub-categories represented by accepted environments.
+## Released Artifacts
+| Artifact | Contents | Use it for |
+| --- | --- | --- |
+| [`Tasks.tar.zst`](https://huggingface.co/datasets/ecnu-icalk/SkillGym/resolve/main/Tasks.tar.zst?download=true) | Published skill-derived task environments, paired variants, and verifiers | Re-running or studying executable tasks |
+| [`skill_library.tar.zst`](https://huggingface.co/datasets/ecnu-icalk/SkillGym/resolve/main/skill_library.tar.zst?download=true) | Human-written skills and supporting assets | Skill analysis or new task construction |
+| [`task_templates.tar.zst`](https://huggingface.co/datasets/ecnu-icalk/SkillGym/resolve/main/task_templates.tar.zst?download=true) | Reusable seed-task templates and fixtures | Extending the task-construction pipeline |
+| [`Trajectories/`](https://huggingface.co/datasets/ecnu-icalk/SkillGym/tree/main/Trajectories) | Eight trajectory collections organized by result type, harness, and teacher model | SFT, trajectory analysis, and procedural-learning research |
+### Release Snapshot
+The large-artifact dataset snapshot is associated with GitHub commit
+[`6ebabba`](https://github.com/ECNU-ICALK/SkillGym/commit/6ebabba).
+The GitHub codebase, dataset repository, and model repository are versioned independently. For exact
+reproduction, record the revision used from each repository.
+## Quick Start
+### Download Trajectory Collections
 ```bash
 python -m pip install -U huggingface_hub datasets
 hf download ecnu-icalk/SkillGym \
+  --include "Trajectories/*.jsonl" \
+  --repo-type dataset \
+  --local-dir .hf/skillgym
+```
+### Stream One Collection with 🤗 Datasets
+```python
+from datasets import load_dataset
+records = load_dataset(
+    "json",
+    data_files={
+        "train": (
+            "hf://datasets/ecnu-icalk/SkillGym/"
+            "Trajectories/"
+            "skill_dependent_claude_code_deepseek_v4_pro.jsonl"
+        )
+    },
+    split="train",
+    streaming=True,
+)
+example = next(iter(records))
+print(example["session_id"])
 ```
+### Download Task-Construction Artifacts
 ```bash
 hf download ecnu-icalk/SkillGym \
+  Tasks.tar.zst skill_library.tar.zst task_templates.tar.zst \
+  --repo-type dataset \
+  --local-dir .hf/skillgym
 ```
+For building new environments, use the
+[SkillGym Task Builder](https://github.com/ECNU-ICALK/SkillGym/tree/main/task_builder).
+## Dataset Organization
+At the release level:
 ```text
+SkillGym/
+├── Tasks.tar.zst
+├── skill_library.tar.zst
+├── task_templates.tar.zst
+└── Trajectories/
+    ├── skill_dependent_*.jsonl
+    └── verifier_passed_*.jsonl
 ```
+The task archives contain the source skill/task relationship, task instructions, environment
+configuration, task-specific verification logic, and paired execution variants used during
+construction and validation.
+Trajectory collections preserve agent interaction traces and verification outcomes for successful
+sampled executions.
+## Terminology
+The same paired skill-availability concept appears under slightly different names across the paper,
+released data, and implementation:
+| Context | Skill available | Skill removed |
+| --- | --- | --- |
+| **Paper / benchmark tables** | `w/ Skills` | `w/o Skills` |
+| **Released dataset variants** | `withskill` | `withoutskill` |
+| **Task Builder internals** | `with_skill` | `no_skill` |
+These names refer to the same paired evaluation concept; the differences reflect presentation,
+released-storage, and implementation conventions.
+## Task Labels and Paired Variants
+SkillGym distinguishes **task-construction labels** from **trajectory-level outcomes**.
+### Skill-Dep.
+A task is labeled **Skill-Dep.** when, under the reference construction configuration:
+1. the environment and verifier are valid;
+2. the paired **with-skill** execution succeeds; and
+3. the paired **without-skill** execution produces a valid reward failure.
+### Verifier-Passed
+A **Verifier-Passed** task has a valid executable environment and outcome verifier but does not satisfy
+the stricter contrastive Skill-Dep. criterion under the reference configuration.
+These labels characterize accepted environments at construction time. They do **not** guarantee that
+every later model or harness will succeed or fail in the same way.
+## Trajectory Data
+Successful trajectories retain the interaction sequence that led to a verifier-confirmed outcome.
+Depending on the source harness, records may include:
+- task and session identifiers;
+- task prompt and skill metadata;
+- environment/runtime metadata;
+- actions and tool calls;
+- observations and tool outputs;
+- final outputs or artifacts;
+- verifier results and execution logs;
+- available diagnostic fields.
+### Schema Note
+The eight trajectory files preserve source-specific trace structures. Some nested fields differ across
+harness/model collections, so the release should **not** be assumed to have one perfectly uniform
+nested JSON schema across every file.
+For the most robust loading behavior:
+- load one trajectory collection at a time; or
+- explicitly normalize compatible files before concatenating them.
+Using an explicit `data_files` path, as in the streaming example above, is preferable to treating every
+trajectory JSONL file as a single homogeneous split.
+## Trajectory Statistics
+| Statistic | Value |
+| --- | ---: |
+| Successful trajectories from Skill-Dep. tasks | **3,722** |
+| Successful trajectories from Verifier-Passed tasks | **4,642** |
+| Explicit target-skill invocation | **3,077 (36.8%)** |
+| Average tool calls per successful trajectory | **49.0** |
+| Average logged text tokens | **63.4k** |
+| Average interaction steps | **35.2** |
+| Corpus-wide maximum tool calls | **350** |
+| Corpus-wide maximum logged text tokens | **342.9k** |
+| Corpus-wide maximum interaction steps | **318** |
+The maxima above are corpus-wide maxima for their respective measures and need not come from the same
+trajectory. Logged/saved text-token counts are dataset statistics and should not be interpreted as API
+billing usage or as a single-request context length.
+## Collection Configurations
+Successful trajectories were collected from multiple harness–model configurations:
+| Harness | Teacher model | Successful trajectories |
+| --- | --- | ---: |
+| Claude Code | DeepSeek V4 Pro | 1,722 |
+| Claude Code | GLM-5.2 | 1,769 |
+| Codex | GPT-5.4 | 1,967 |
+| Codex | Nex-N2-Pro | 2,906 |
+This diversity broadens the interaction traces beyond one agent configuration, but the released
+trajectory distribution should not be treated as representative of every model, harness, task domain,
+or deployment setting.
+## Intended Uses
+SkillGym can support research on:
+- supervised fine-tuning from verified agent trajectories;
+- agentic problem solving and long-horizon workflow learning;
+- skill-conditioned versus skill-free execution;
+- executable-environment and verifier-based evaluation;
+- outcome-based reinforcement learning;
+- task-generation and verifier design;
+- trajectory analysis and procedural failure modes.
+## Limitations
+- Skill-Dep. and Verifier-Passed are reference-configuration labels, not universal properties.
+- Category coverage is broad but uneven.
+- Trajectory collections are generated by a finite set of harness–model configurations.
+- Successful trajectories are a selected subset of sampled executions and should not be interpreted
+  as an unbiased behavioral distribution.
+- Some trajectory files preserve harness-specific schema differences.
+- Source skills, fixtures, and supporting assets may carry upstream notices or additional licensing
+  requirements.
+## Related Resources
+| Resource | Scope |
+| --- | --- |
+| [GitHub / SkillGym](https://github.com/ECNU-ICALK/SkillGym) | Task Builder, construction pipeline, documentation, and paper materials |
+| [SkillGym-Agent](https://huggingface.co/ecnu-icalk/SkillGym-Agent) | Released **All Teachers** checkpoint trained on successful SkillGym trajectories |
+| [Task Builder guide](https://github.com/ECNU-ICALK/SkillGym/blob/main/task_builder/README.md) | Building and validating new skill-grounded environments |
+| [Paper](https://github.com/ECNU-ICALK/SkillGym/blob/main/assets/Internalizing_Large_Scale_Human_Written_Skills_into_LLMs_for_Real_World_Problem_Solving.pdf) | Full method, dataset analysis, experiments, and ablations |
 ## Citation
+The accompanying paper is currently under double-blind review. Until final publication metadata is
+available, please use:
 ```bibtex
+@misc{skillgym2026,
+  title = {Internalizing Large-Scale Human-Written Skills into LLMs for Real-World Problem Solving},
   year  = {2026},
   url   = {https://github.com/ECNU-ICALK/SkillGym}
 }
 ```
+## License
+The dataset is marked **`license: other`** because it aggregates artifacts that may retain distinct
+upstream notices or additional terms.
+Repository-level SkillGym code is released under the **MIT License**. The SkillGym-Agent checkpoint
+is released separately under **Apache-2.0**.
+When redistributing source skills, fixtures, or supporting assets, preserve the corresponding upstream
+notices and licensing terms.