add clawgym-filtered: jsonl + slime/torchtitan parquet + torchtitan_sft arrow

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+clawgym.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,124 @@
+---
+license: apache-2.0
+language:
+- en
+task_categories:
+- text-generation
+tags:
+- swe
+- agentic
+- tool-calling
+- openclaw
+- sft
+size_categories:
+- 10K<n<100K
+---
+
+# ClawGym-Filtered
+
+OpenClaw agentic trajectories (MiniMax-M2.5 + GLM-5.1) filtered and reformatted
+for SFT / mid-training of agentic coding models.
+
+## Summary
+
+- **Rows**: 15,777
+- **Tokens**: 342M (Qwen3-4B tokenizer, training chat template)
+- **Source**: [RUC-AIBOX/ClawGym-Trajectory](https://huggingface.co/datasets/RUC-AIBOX/ClawGym-Trajectory)
+- **Scaffold**: OpenClaw (19-tool toolset)
+- **Generators**: MiniMax-M2.5, GLM-5.1
+
+## Pipeline
+
+24,497 raw → 15,777 final (64.4% retention):
+
+| Stage | Out | Δ | Reason |
+|---|---:|---:|---|
+| raw (HF) | 24,497 | — | RUC-AIBOX/ClawGym-Trajectory |
+| 07 schema recovery | 19,863 | -4,634 | residual tool_call schema violations |
+| 08 preprocessed | 19,185 | -678 | empty_asst / identity / leak / refusal |
+| 10 CJK filter | 18,010 | -1,175 | Chinese leakage |
+| 11 hacking gate | 18,009 | -1 | skip mode, edge case |
+| 12 quality judge | 15,791 | -2,218 | Qwen3.6-35B 5-rubric individual mode |
+| 15 postprocess | **15,777** | -14 | min_tokens / compression / leak |
+
+## Reasoning-content shift
+
+ClawGym models emit chain-of-thought as the prose body of an assistant turn
+that also carries `tool_calls`. We move that text into `reasoning_content`
+and null out `content`, so only the **final no-tool-call assistant turn**
+keeps user-visible content. Pure tool-call turns (no prose, no reasoning)
+are left untouched.
+
+Verified pre-shift: every row has exactly one no-tool-call assistant turn,
+always last (0 mid-trajectory cases across 15,777 rows).
+
+| Asst turns | Count | After shift |
+|---|---:|---|
+| with tool_calls, non-empty content | 114,891 | content → reasoning_content, content = null |
+| with tool_calls, empty content | 21,478 | unchanged |
+| no tool_calls (final reply) | 15,777 | unchanged, content kept |
+
+## Files
+
+| Path | Format | Size |
+|---|---|---:|
+| `clawgym.jsonl` | raw final JSONL (one `{context, meta, num_tokens}` per line) | 1.35 GB |
+| `slime/clawgym.parquet` | OpenAI-spec 6-column, messages/tools as JSON string | 0.44 GB |
+| `torchtitan/clawgym.parquet` | 4-column, `context` as native `list<struct>` | 0.40 GB |
+| `torchtitan_sft/` | Arrow dataset, 3 shards (256K max-token cap) | 1.30 GB |
+
+### Slime parquet schema
+
+```
+instance_id (str) | domain (str) | source (str) | messages (str) | tools (str) | num_tokens (int64)
+```
+
+`messages` / `tools` are JSON strings. Assistant turn order is fixed:
+`{role, content, reasoning_content, tool_calls}` (tool_calls = [] when empty).
+
+### Torchtitan parquet schema
+
+```
+uuid_or_instance_id (str) | context (list<struct>) | tools (str) | meta (str)
+```
+
+`context` is a native `list<struct>` (no json.loads needed for the outer
+list). Per-turn struct (alphabetical): `{content, reasoning, role,
+tool_call_id, tool_calls}`. The only inner JSON string is
+`context[i].tool_calls[j].function.arguments`.
+
+### JSONL schema
+
+```jsonc
+{
+  "context": [
+    {"role": "system",    "content": "..."},
+    {"role": "user",      "content": "..."},
+    {"role": "assistant",
+     "content": null,
+     "reasoning_content": "Let me read all the input files first.",
+     "tool_calls": [{"id": "...", "type": "function",
+                     "function": {"name": "read", "arguments": "{\"path\":\"...\"}"}}]},
+    {"role": "tool",      "tool_call_id": "...", "content": "..."},
+    {"role": "assistant", "content": "Done. Here is the result..."}
+  ],
+  "meta": {"source": "clawgym", "instance_id": "clawgym::row_0", "tools": [...]},
+  "num_tokens": 23088
+}
+```
+
+## Loading
+
+```python
+# Arrow dataset (torchtitan_sft)
+from datasets import load_dataset
+ds = load_dataset("hyunwoongko/clawgym-filtered", data_dir="torchtitan_sft")
+
+# Parquet
+import pyarrow.parquet as pq
+table = pq.read_table("torchtitan/clawgym.parquet")  # or "slime/clawgym.parquet"
+```
+
+## License
+
+Apache 2.0 (inherited from upstream RUC-AIBOX/ClawGym-Trajectory).

clawgym.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:003e3252616b6d2b93cfc312d6001426f9825c8cb500d99d96de5e6c63faf6ab
+size 1352719151

slime/clawgym.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dac5e63f980ab47a9d12bdce18fdf8e19fe7e9433af4cb86566e7ffc1ffe4770
+size 444780610

torchtitan/clawgym.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:935031c79c6431ce88785d234fcfbfb2d5cc84d17284c41b6e702583eb4c2fcd
+size 401084564

torchtitan_sft/data-00000-of-00003.arrow

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e2e08e2cc9f4593684af3368c4e6cc5b53efc195d3d5857b96d7bb21500a37d9
+size 450796944

torchtitan_sft/data-00001-of-00003.arrow

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2a109af00aee17397ca5fb84f32bace8febe0a217a8b8879d7d089cec9c43ff
+size 443969808

torchtitan_sft/data-00002-of-00003.arrow

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d248b6dc2169a3785461c421d281beea44e4942056b1fdcdc737a9194fa228c
+size 403818184

torchtitan_sft/dataset_info.json

@@ -0,0 +1,60 @@
+{
+  "citation": "",
+  "description": "",
+  "features": {
+    "context": {
+      "feature": {
+        "content": {
+          "dtype": "string",
+          "_type": "Value"
+        },
+        "function_name": {
+          "dtype": "string",
+          "_type": "Value"
+        },
+        "reasoning": {
+          "dtype": "string",
+          "_type": "Value"
+        },
+        "role": {
+          "dtype": "string",
+          "_type": "Value"
+        },
+        "tool_call_id": {
+          "dtype": "string",
+          "_type": "Value"
+        },
+        "tool_calls": {
+          "feature": {
+            "function": {
+              "arguments": {
+                "dtype": "string",
+                "_type": "Value"
+              },
+              "name": {
+                "dtype": "string",
+                "_type": "Value"
+              }
+            },
+            "id": {
+              "dtype": "string",
+              "_type": "Value"
+            },
+            "type": {
+              "dtype": "string",
+              "_type": "Value"
+            }
+          },
+          "_type": "List"
+        }
+      },
+      "_type": "List"
+    },
+    "tools": {
+      "dtype": "string",
+      "_type": "Value"
+    }
+  },
+  "homepage": "",
+  "license": ""
+}

torchtitan_sft/state.json

@@ -0,0 +1,19 @@
+{
+  "_data_files": [
+    {
+      "filename": "data-00000-of-00003.arrow"
+    },
+    {
+      "filename": "data-00001-of-00003.arrow"
+    },
+    {
+      "filename": "data-00002-of-00003.arrow"
+    }
+  ],
+  "_fingerprint": "09e42c6c7a952123",
+  "_format_columns": null,
+  "_format_kwargs": {},
+  "_format_type": null,
+  "_output_all_columns": false,
+  "_split": null
+}