Add normalized multimodal interactions and preserved image assets

README.md

@@ -1,37 +1,111 @@
 ---
-pretty_name: pi-embodied Robotics R&D Interactions
+pretty_name: pi-embodied Robotics R&D Multimodal Interactions
 language:
 - en
 - zh
 configs:
 - config_name: default
   data_files:
-  - split: staging
-    path:
-    - data/batches/*/*.jsonl.gz
-    - data/batches/*/*.jsonl
+  - split: train
+    path: data/train-*.parquet
+dataset_info:
+  features:
+  - name: id
+    dtype: string
+  - name: schema_version
+    dtype: string
+  - name: trajectory_unit
+    dtype: string
+  - name: source
+    dtype: string
+  - name: session_id
+    dtype: string
+  - name: turn_id
+    dtype: string
+  - name: model
+    dtype: string
+  - name: messages
+    list:
+    - name: role
+      dtype: string
+    - name: content
+      list:
+      - name: type
+        dtype: string
+      - name: text
+        dtype: string
+      - name: path
+        dtype: string
+      - name: mime_type
+        dtype: string
+    - name: tool_calls
+      list:
+      - name: id
+        dtype: string
+      - name: type
+        dtype: string
+      - name: name
+        dtype: string
+      - name: arguments
+        dtype: string
+    - name: tool_call_id
+      dtype: string
+    - name: name
+      dtype: string
+    - name: timestamp
+      dtype: string
+  - name: messages_json
+    dtype: string
+  - name: tools_json
+    dtype: string
+  - name: tools_status
+    dtype: string
+  - name: images
+    list: image
+  - name: media_refs
+    list: string
+  - name: media_status
+    dtype: string
+  - name: outcome
+    dtype: string
+  - name: repository_url
+    dtype: string
+  - name: base_commit
+    dtype: string
+  - name: patch
+    dtype: string
+  - name: test_result_json
+    dtype: string
+  - name: provenance_json
+    dtype: string
+  - name: native_records_json
+    dtype: string
+  - name: lerobot_status
+    dtype: string
 ---
 
-# pi-embodied Robotics R&D Interactions
+# pi-embodied Robotics R&D Multimodal Interactions
 
-Project-scoped Codex and Claude coding/research interactions for pi-embodied, plus retained related relay requests. Task families have not yet been curated for model training.
+This public dataset contains 2,396 project-scoped agent turns, including 183 turns with recoverable, published images. It preserves successful, unsuccessful, interrupted and unverified interactions without inventing outcome labels.
 
-## Collection
+## Storage
 
-The first batch is dated 2026-10-06. Sources include local Codex and Claude sessions, related requests retained by the audited relay on sg-ai-gateway, and project-scoped Codex, Claude and pi agent sessions from AutoDL, including native pi agent runtime logs. Runtime collection excludes dependency/test fixtures, standalone recipes and dataset metadata. Selection uses explicit pi-embodied references or a project-specific working directory. Unrelated conversation turns, system/developer instructions, binary media, and encrypted reasoning are omitted. Relay responses are retained only when their current conversation block matches project scope.
+- `data/train-*.parquet`: normalized interactions with an explicit Arrow schema and Zstandard compression.
+- `media/<hash-prefix>/<sha256>.<extension>`: deduplicated original image bytes, referenced by messages and exposed through the Hugging Face `images` feature.
+- `manifest.json`: row counts, media recovery status and image checksums.
+- `data/batches/` and `manifests/`: immutable historical text archives, excluded from the default dataset configuration.
+- `load_multimodal.py`: loading helpers for interactions, chat messages and execution exports.
 
-## Storage and provenance
+## Conversation schema
 
-This repository is public. JSONL payloads, compressed where practical, live under `data/batches/<batch-id>/`; manifests record source counts, redactions, byte sizes and SHA-256 checksums. Each row records its source, model when available, source session/trace identity, selection rule, schema version and stable content hash. Sources remain distinguishable for later deduplication.
+Each row corresponds to a previously selected task turn, not a certified full robot episode. `messages` stores ordered user/assistant/tool messages, content parts, call IDs, function names, arguments and timestamps. `messages_json` stores an OpenAI-style conversation with `tool_calls.function` and `image_url` content parts. Decode it with `json.loads`; image URLs resolve to this repository. Visible reasoning summaries are retained when available; encrypted provider content is excluded.
 
-Local and AutoDL rows contain `records`; relay rows contain `payload.request_messages` and an optional `payload.response_body`. Tool arguments and results retain their native text structure. The schemas are intentionally marked as staging rather than a unified training format.
+`images` is a list of Hugging Face Image features; `media_refs` identifies the corresponding repository-relative files. Original source records, after media extraction and text redaction, are available in `native_records_json`. `provenance_json` links every normalized row to its immutable archive identity. The same row IDs and source classification are retained.
 
-## Data handling
+`tools_json` is `[]` and `tools_status` is `not_recorded` where the original collection did not preserve tool definitions. Tool definitions are not reconstructed from argument examples. Repository revisions, patches, test results and task outcomes remain null or unknown when not independently linked to the recorded turn.
 
-Automatic redaction removes recognizable tokens, private keys, authorization headers, secret assignments, known credentials, email addresses and IP literals. Personal home-directory components are normalized. Automatic redaction is best effort and requires further review before training. Overlapping histories and cross-source duplicates require additional curation. A source inspection or tool result does not establish task success.
+## Media recovery and handling
 
-No public redistribution or training license is assigned in this initial collection. Original third-party content retains its own applicable rights.
+Original inline images were recovered from source logs for previously selected turns. Recoverable explicit local image attachments are included. Referenced files that no longer exist remain marked as missing. Remote URLs are not crawled to reconstruct unavailable content. Text is filtered for recognizable credentials and personal identifiers; image OCR withholds recognizable sensitive content for private review. OCR is a filter rather than a guarantee. Original source logs and withheld images remain on their private source/staging storage.
 
-## Future batches
-
-Append immutable batches and update their manifests. Do not overwrite earlier payloads. Future policy collections should preserve episode boundaries, observation/action alignment, environment identity, task outcomes and timestamps. Future R&D collections should link coding interactions with repository revisions and independently verified tests/results.
+The default `train` split denotes a usable storage split, not a claim that all examples are suitable for supervised training. Outcome-based curation, overlapping-history deduplication, context completion and tool-schema capture remain necessary for individual training recipes. No new redistribution or training license is assigned; original third-party material retains its applicable rights.

data/train-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c4122b9c643e56805e87b25f0a5316e653c3a396747aaf8a0c0af24eec2b760
+size 203055525

load_multimodal.py

@@ -0,0 +1,26 @@
+#!/usr/bin/env python3
+"""Load normalized interactions and individual LeRobot exports from the Hub."""
+import json
+import pathlib
+from datasets import load_dataset
+from huggingface_hub import snapshot_download
+
+
+def load_interactions(stream='policy', split='train'):
+    return load_dataset(f'LV-Robotics-Lab/pi-embodied-{stream}-interactions', split=split)
+
+
+def as_chat(row):
+    """OpenAI-style messages, including tool calls and remote image URLs."""
+    return json.loads(row['messages_json'])
+
+
+def load_execution_export(name, revision='main'):
+    from lerobot.datasets.lerobot_dataset import LeRobotDataset
+    repo = 'LV-Robotics-Lab/pi-embodied-policy-interactions'
+    snapshot = pathlib.Path(snapshot_download(repo, repo_type='dataset', revision=revision,
+                                              allow_patterns=[f'lerobot/{name}/**']))
+    root = snapshot/'lerobot'/name
+    if not (root/'meta/info.json').exists():
+        raise FileNotFoundError(f'Unknown execution export: {name}')
+    return LeRobotDataset(repo_id=repo, root=root, download_videos=False)