v0.3 quality hardening: SFT-ready schema, portable media manifests, and seven-config parquet release

README.md

@@ -1,72 +1,83 @@
 ---
-pretty_name: Streaming Omni Datasets
+pretty_name: StreamingOmniDatasets v0.3 Quality Hardened
+license: other
 task_categories:
-- visual-question-answering
-- video-text-to-text
-- audio-classification
 - automatic-speech-recognition
+- audio-classification
+- video-text-to-text
+- visual-question-answering
 - text-to-speech
-license: other
+configs:
+- config_name: StreamingAudioInteraction
+  data_files:
+  - split: train
+    path: data/StreamingAudioInteraction/train-00000-of-00001.parquet
+  - split: eval
+    path: data/StreamingAudioInteraction/eval-00000-of-00001.parquet
+- config_name: StreamingASRTranslation
+  data_files:
+  - split: train
+    path: data/StreamingASRTranslation/train-00000-of-00001.parquet
+  - split: eval
+    path: data/StreamingASRTranslation/eval-00000-of-00001.parquet
+- config_name: SpeechOutputAlignment
+  data_files:
+  - split: train
+    path: data/SpeechOutputAlignment/train-00000-of-00001.parquet
+  - split: eval
+    path: data/SpeechOutputAlignment/eval-00000-of-00001.parquet
+- config_name: LiveAVCommentary
+  data_files:
+  - split: train
+    path: data/LiveAVCommentary/train-00000-of-00001.parquet
+  - split: eval
+    path: data/LiveAVCommentary/eval-00000-of-00001.parquet
+- config_name: AVJointReasoning
+  data_files:
+  - split: train
+    path: data/AVJointReasoning/train-00000-of-00001.parquet
+  - split: eval
+    path: data/AVJointReasoning/eval-00000-of-00001.parquet
+- config_name: ProactiveAVInteraction
+  data_files:
+  - split: train
+    path: data/ProactiveAVInteraction/train-00000-of-00001.parquet
+  - split: eval
+    path: data/ProactiveAVInteraction/eval-00000-of-00001.parquet
+- config_name: OCRScreenGrounding
+  data_files:
+  - split: train
+    path: data/OCRScreenGrounding/train-00000-of-00001.parquet
+  - split: eval
+    path: data/OCRScreenGrounding/eval-00000-of-00001.parquet
 ---
 
-# Streaming Omni Datasets
-
-Repository target: `skyzhou06/StreamingOmniDatasets`
-
-This package contains Phase 1 and Phase 2 local outputs for Streaming Omni dataset construction.
-
-## Dataset List
-
-| Dataset | Train | Eval | High quality |
-| --- | ---: | ---: | ---: |
-| StreamingAudioInteraction | 950 | 50 | 1000 |
-| StreamingASRTranslation | 950 | 50 | 1000 |
-| SpeechOutputAlignment | 950 | 50 | 1000 |
-| LiveAVCommentary | 950 | 50 | 1000 |
-| AVJointReasoning | 950 | 50 | 1000 |
-
-## Schema
-
-Rows follow the repository `configs/schema.json` timeline schema with audio, video, text, evidence, and quality fields.
-
-## Construction Pipeline
-
-Phase 1 builds streaming audio interaction, ASR/translation, and speech output alignment. Phase 2 adds live AV commentary and AV joint reasoning from non-benchmark local or public/open online media.
-
-## Online Sources
-
-- Successfully used source distribution: `{'internet_archive_public': 4}`
-- Usable downloaded/public clips: `4`
-- Download failures by reason: `{}`
-
-## Tool Usage Transparency
-
-Reports record whether WhisperX, CosyVoice, VLM captioning, and audio tagging used real models, remote execution, or deterministic fallback mode. Phase 2 in this package uses real downloaded media with fallback caption/tag labels because no real VLM/audio tagger backend was configured locally or remotely.
-
-## Benchmark Exclusion
-
-Benchmark datasets listed in the project reports are not used as training sources.
-
-## Limitations
-
-The public-media source subset is intentionally small. Fallback captions/tags are conservative labels grounded in source metadata and sampled frames/audio, not real VLM or CLAP/BEATs inference. Source license notes are preserved and should be audited before downstream redistribution decisions.
-
-## Intended Use
-
-Research and development of streaming multimodal dataset construction pipelines.
+# StreamingOmniDatasets v0.3 Quality Hardened
 
-## Not Intended For
+Streaming multimodal SFT data with seven configurations, a canonical `streaming_omni_v0.3.1` schema, 950 train rows and 50 eval rows per configuration.
 
-Production safety-critical deployment or claims of benchmark performance.
+Raw media is not universally redistributed. Each row declares one `metadata.media_access` strategy: upstream reconstruction for restricted Seamless Interaction derivatives, or deterministic generation for synthetic speech/OCR assets. See [docs/MEDIA_RECONSTRUCTION.md](docs/MEDIA_RECONSTRUCTION.md).
 
-## Citation
+## Loading
 
-Citation placeholder.
+```python
+from datasets import load_dataset
 
-## License
+ds = load_dataset(
+    "skyzhou06/StreamingOmniDatasets",
+    "ProactiveAVInteraction",
+    split="train",
+)
+print(ds[0]["timeline"], ds[0]["interaction_policy"])
+```
 
-License placeholder; source-specific license notes are preserved in row metadata and reports.
+## Integrity and limitations
 
-## Phase Status
+- Training lineage excludes the benchmark banlist documented in `docs/SOURCE_PROVENANCE.md`.
+- Seamless Interaction-derived data retains the CC BY-NC 4.0 non-commercial restriction; the repository uses `license: other` because sources differ.
+- Qwen2.5-VL annotations are accepted only after direct or safely repaired JSON parsing and schema validation; final statistics are in `docs/V0_3_RELEASE_NOTES.md`.
+- WhisperX timestamps are segment-level, not word-level.
+- OCR assets in this hardened release are deterministic generated examples with generator ground truth; they are not represented as real OCR-model detections or verified public-sign photographs.
+- Agent review is not human review. Human-review CSV/JSONL artifacts are provided separately in the project outputs.
 
-See `reports/PHASE1_REPORT.md`, `reports/PHASE2_REPORT.md`, and `reports/validation_report.json`.
+See `docs/SCHEMA.md`, `docs/INTERNAL_SFT_USAGE.md`, `docs/LICENSE_AND_PROVENANCE_MATRIX.md`, and `docs/LIMITATIONS.md`.

data/AVJointReasoning/{eval.parquet → eval-00000-of-00001.parquet}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ee6b85cd6911f61af187385347b1eb8b35951897f86bec3f163aa1fcd8cfbef8
-size 23085
+oid sha256:eec74340c0a3ca9204a5f14520cea5c4beb60fd6d3f232503d29e316154c2735
+size 23041

data/AVJointReasoning/{high_quality.parquet → train-00000-of-00001.parquet}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1dff16ab2a01ece46aa8694c8716e086c0cb8e4eba3794f53cabda767909f808
-size 32014
+oid sha256:5e795ba2534f78b0eadc89f5181ff67d4f189441b0b73a961957e69686fdf34d
+size 47325

data/LiveAVCommentary/{eval.parquet → eval-00000-of-00001.parquet}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7fa635d4a2779e4c286bfa09b6f5cfceb5fac0038edf44adae18448506e91e44
-size 21383
+oid sha256:3302083d14dd29c7dd49726747ac5141223ec582b0575d7cb68ae4984c0cccbe
+size 21690

data/LiveAVCommentary/high_quality.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:ee3d2df9a36d79802b97be96cc21bf083a2b644f3ff98c0a9c3aaa052d43c6cf
-size 30423

data/{AVJointReasoning/train.parquet → LiveAVCommentary/train-00000-of-00001.parquet}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6b060f9ba9ad62fd5fb42a3ec2d2d1f2b94643a0f02a1a7b97db8758c7379bbe
-size 31646
+oid sha256:8dcbb9ba0fab7f05fd25342ab7b4f7412d0354cd156a7e4c9379f961edcdab3b
+size 47525

data/LiveAVCommentary/train.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:15569dc44dfb4dfb2aaa3d2311b8bcadab4997db8d1b57dae33ab7fa35d14aaa
-size 30055

data/OCRScreenGrounding/eval-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60d1f7cd78cadc354cf86d00f7ffa8e7c4ba8729881a755cf888f6785f7f3894
+size 34311

data/OCRScreenGrounding/train-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ee38d5838da41062f12889d1ce7d84d213cbe6beae9bf471627462ee47c2548e
+size 102813

data/ProactiveAVInteraction/eval-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7b13344eab735d5b92106609032aabd42b4eb854175ca78d11f3ab5978e58ef3
+size 38158

data/ProactiveAVInteraction/train-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff4c474d920e5bd870eeafc865b3d9377254ebd6edd6cf24b9a82a73d3865ed6
+size 95658

data/SpeechOutputAlignment/eval-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f011da08a4e10649d251c3b7095c806055fc4c2432a05cb7cd1cdab1d183e2f9
+size 26673

data/SpeechOutputAlignment/eval.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:402b29b0e74d80ab77518421c6abddd0c790b41fa3549fdc1dbb8bc91b452236
-size 20074

data/SpeechOutputAlignment/high_quality.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:21d2a15e81458816ddc8828a7ee192a4f7a094fcd34619cb6e409683ed35bef8
-size 58236

data/SpeechOutputAlignment/train-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6da8a4bc6c1eb88f549b8733e836d7c069971b7a42a459fb953d202c1c545405
+size 73227

data/SpeechOutputAlignment/train.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:e6e0cbb971fb670c91c9841e430d726a4e8cba00544350306caa80645dc70e89
-size 56535

data/StreamingASRTranslation/eval-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d2d569ac34ef668ddcb9eb168048734e727146ee8a9f311b59a50cac50ebf2b0
+size 24634

data/StreamingASRTranslation/eval.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:f7aef5eacb5ec8dfde763974c3af0293a3dcdf2edabb7e3025ae0ed8a27c36c8
-size 17486

data/StreamingASRTranslation/high_quality.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:6e7912c4d5043da75a824487f32441d65ef69a096651395c74e7b03e21b65644
-size 44128

data/StreamingASRTranslation/train-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d028577a571f044e23a31ab54afc7f12ad8e348cfd43278830d35bfcff039859
+size 72350

data/StreamingASRTranslation/train.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:73bf33409be955b896e6022277e4a52e8334114115cff6a21cf128ff60eb28d9
-size 42940

data/StreamingAudioInteraction/eval-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45bc4bb71400d4cd1aa969d871e1cece4e521446334970f7c28d7842e17f5e50
+size 23006

data/StreamingAudioInteraction/eval.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:0cc72e9acd98d25f5644ea401c9d577121f0b058f43591b7a24bb8bcf6732038
-size 16404

data/StreamingAudioInteraction/high_quality.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:00024ea6cca34eec023ad785c3576f57f0fd7ec67c4ac2b391847e7d9d147f61
-size 43516

data/StreamingAudioInteraction/train-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a44dac6c670aac24713b9373c953ce9f652898ab098f60f20c6ea89ee346b883
+size 58771

data/StreamingAudioInteraction/train.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:71aa708acfcdf597a3192b7b0b82a1c161b777600d03b5bb515d073e347eec9e
-size 42171

dataset_infos.json

@@ -1,34 +0,0 @@
-{
-  "repo_id": "skyzhou06/StreamingOmniDatasets",
-  "datasets": {
-    "StreamingAudioInteraction": {
-      "train": 950,
-      "eval": 50,
-      "high_quality": 1000
-    },
-    "StreamingASRTranslation": {
-      "train": 950,
-      "eval": 50,
-      "high_quality": 1000
-    },
-    "SpeechOutputAlignment": {
-      "train": 950,
-      "eval": 50,
-      "high_quality": 1000
-    },
-    "LiveAVCommentary": {
-      "train": 950,
-      "eval": 50,
-      "high_quality": 1000
-    },
-    "AVJointReasoning": {
-      "train": 950,
-      "eval": 50,
-      "high_quality": 1000
-    }
-  },
-  "include_phases": [
-    "phase1",
-    "phase2"
-  ]
-}

docs/INTERNAL_SFT_USAGE.md

@@ -0,0 +1,7 @@
+# Internal SFT Usage
+
+`outputs/internal_sft_v0_3/` contains fourteen Parquet splits, materialized package-relative media, and per-config SFT JSONL.
+
+Explicit adapters are provided for generic streaming, Qwen-Omni-style messages, and MiniCPM-o-style conversations. They intentionally do not claim tokenizer, processor, audio-codec, or speech-decoder compatibility across all model releases.
+
+Run `scripts/33_validate_sft_readiness.py`, `scripts/34_sft_dataloader_smoke_test.py`, and `scripts/34b_mini_sft_smoke_test.py` before training.

docs/LICENSE_AND_PROVENANCE_MATRIX.md

@@ -0,0 +1,13 @@
+# License and Provenance Matrix
+
+| Configuration | Source | Split | Media license | Annotation/license note | Raw media | Recovery | Restriction |
+|---|---|---|---|---|---|---|---|
+| StreamingAudioInteraction | existing text supervision + generated Piper speech | train/eval | generated asset | project annotation license not separately asserted | generated internally, omitted publicly | deterministic generation | do not infer a permissive upstream license |
+| StreamingASRTranslation | existing transcripts + generated Piper speech | train/eval | generated asset | project annotation license not separately asserted | generated internally, omitted publicly | deterministic generation | do not infer a permissive upstream license |
+| SpeechOutputAlignment | existing response text + generated Piper speech | train/eval | generated asset | project annotation license not separately asserted | generated internally, omitted publicly | deterministic generation | do not infer a permissive upstream license |
+| LiveAVCommentary | facebook/seamless-interaction | train | CC BY-NC 4.0 | derived annotations | omitted | upstream manifest | attribution, non-commercial |
+| AVJointReasoning | facebook/seamless-interaction | train | CC BY-NC 4.0 | derived annotations | omitted | upstream manifest | attribution, non-commercial |
+| ProactiveAVInteraction | facebook/seamless-interaction | train | CC BY-NC 4.0 | derived Qwen/CLAP/WhisperX annotations | omitted | upstream manifest + clip boundaries | attribution, non-commercial |
+| OCRScreenGrounding | deterministic generated UI/document/table/terminal/signage | generated | generated asset | generator ground truth | generated internally, omitted publicly | deterministic generation | not presented as public-sign media or OCR-model output |
+
+The repository uses `license: other`; no single permissive license is asserted for all configurations.

docs/LIMITATIONS.md

@@ -0,0 +1,10 @@
+# Limitations
+
+- Seamless Interaction-derived media is non-commercial and is reconstructed from upstream rather than redistributed.
+- Upstream archives or mirrors may become unavailable.
+- Generated speech is useful for serialization and pipeline training checks but is not a substitute for diverse natural speech.
+- OCR examples are generated and do not measure real-world OCR robustness.
+- VLM/CLAP/WhisperX annotations can contain model errors even after structural validation.
+- WhisperX timing is segment-level, not word-level.
+- The minimal 12-step training test validates data/collator/backprop integration only; it is not evidence of compatibility with every production model.
+- Agent review is not human review.

docs/MEDIA_RECONSTRUCTION.md

@@ -0,0 +1,17 @@
+# Media Reconstruction
+
+Each row points to a configuration manifest through `metadata.media_access.manifest_id`.
+
+```bash
+python scripts/31_reconstruct_media.py \
+  --config ProactiveAVInteraction \
+  --output-dir outputs/reconstructed_media \
+  --limit 10
+```
+
+Modes:
+
+- `reconstruct`: download the declared upstream archive, extract its exact member, and apply the recorded clip boundaries. CC BY-NC media is not included in the repository.
+- `generated`: render OCR text from the deterministic recipe or synthesize speech with the documented Piper voice.
+
+The public package contains no required server-local path. Network availability and upstream license compliance remain the user’s responsibility.

docs/MODEL_BACKENDS.md

@@ -0,0 +1,7 @@
+# Model Backends
+
+- Qwen2.5-VL-7B-Instruct: real frame-conditioned invocations. Only direct or safely repaired structured JSON that passes field validation is accepted.
+- CLAP: real audio-tagging scores retained for ProactiveAVInteraction.
+- WhisperX small: real segment transcription. Empty no-speech results are explicitly marked `asr_status=no_speech`; timestamps are segment-level.
+- OCRScreenGrounding: deterministic generator ground truth, not OCR-model inference.
+- Piper `en_US-lessac-medium`: deterministic materialization for the internal speech package. Generated speech is documented as synthetic.

docs/RELEASES.md

@@ -0,0 +1,11 @@
+# Releases
+
+## v0.3-quality-hardened
+
+- Seven configurations, each 950 train / 50 eval.
+- Canonical `streaming_omni_v0.3.1` schema and deterministic timeline chunk IDs.
+- Proactive evidence IDs and OCR grounding/target resolution.
+- Separate materialized internal SFT and manifest-backed public packages.
+- Fourteen consistent Parquet split files and explicit Viewer configuration.
+
+Earlier v0.3 candidates and the previous remote layout are not publication-quality because their package, media, schema, and Viewer gates were incomplete.

docs/SCHEMA.md

@@ -0,0 +1,19 @@
+# Canonical Schema `streaming_omni_v0.3.1`
+
+Every row contains `schema_version`, a unique `id`, config/split/source lineage, modalities, duration, a non-empty timeline, evidence, quality metadata, and `metadata.media_access`.
+
+Every timeline chunk has a deterministic `chunk_0000`-style ID, monotonic `t_start < t_end`, streaming ASR/OCR/audio/visual events, control state, and optional target deltas. Chunk times stay inside `duration_sec`.
+
+`metadata.media_access` declares exactly one public strategy: `reconstruct` with a stable upstream URL and manifest entry, or `generated` with a deterministic recipe. The internal package rewrites it to `included` with a package-relative path.
+
+## ProactiveAVInteraction
+
+`interaction_policy` includes `should_speak`, `speak_timing`, `trigger_type`, `reason_for_silence`, resolvable `evidence_chunk_ids`, optional latency, priority, and interruption permission. A `genuinely_joint` label is used only when both independent audio and visual evidence are materially required.
+
+## OCRScreenGrounding
+
+`ocr_grounding` includes `required`, non-empty `text_instances`, normalized `[x1,y1,x2,y2]` boxes, frame timestamps, confidence, source, resolvable `target_instance_ids`, and temporal state fields. Generated examples use `source=generator_ground_truth`; no generator truth is presented as OCR-model inference.
+
+## Model mode and quality
+
+`metadata.model_mode` records backend modes. `quality.status` must be `pass` for release rows. WhisperX timestamps are segment-level.

docs/SOURCE_PROVENANCE.md

@@ -0,0 +1,9 @@
+# Source Provenance
+
+All rows retain `source_id`, `source_dataset`, original identifiers, split assignment, license notes, and a manifest ID.
+
+- `LiveAVCommentary`, `AVJointReasoning`, and `ProactiveAVInteraction` derive from `facebook/seamless-interaction/train` media under CC BY-NC 4.0. Raw media is omitted; manifests point to upstream archives and clip boundaries.
+- The three speech configurations use deterministic generated speech from their existing text supervision. They do not claim that the speech is naturally recorded.
+- `OCRScreenGrounding` uses deterministic rendered text and generator ground truth. Entries previously called public signage lacked verifiable public-media provenance and were honestly reclassified as generated signage.
+
+Forbidden training sources include LiveSports-3K, Daily-Omni, WorldSense, JointAVBench, AVUT-Human, Video-MME eval, OmniPro, Proactive-Sound-Bench, LongVideoBench, MLVU, Video-Holmes, FutureOmni, SeedTTS eval, LongTTS eval, TextVQA VAL, OCRBench eval, DocVQA eval, HallusionBench, and MMHal eval.

docs/TASK_MAPPING.md

@@ -0,0 +1,13 @@
+# Task Mapping
+
+| Configuration | Primary supervision | Media strategy |
+|---|---|---|
+| StreamingAudioInteraction | streaming speech interaction/control | generated speech |
+| StreamingASRTranslation | streaming ASR/translation text | generated speech |
+| SpeechOutputAlignment | response speech timing/alignment | generated speech |
+| LiveAVCommentary | live audiovisual commentary | upstream reconstruction |
+| AVJointReasoning | grounded audiovisual question answering | upstream reconstruction |
+| ProactiveAVInteraction | speak/wait/never policy and evidence | upstream reconstruction |
+| OCRScreenGrounding | OCR transcription, instances, boxes, temporal state | deterministic generated image |
+
+Benchmark names may inform taxonomy and metrics only; the banlist in `SOURCE_PROVENANCE.md` is excluded from row lineage, source manifests, and prompt inputs.

docs/V0_2_1_RELEASE_NOTES.md

@@ -0,0 +1,25 @@
+# v0.2.1 Release Notes
+
+## What Changed
+
+- **No parquet/data changes**: All v0.2 parquet data is unchanged.
+- **No v0.2 data rebuild**: Scripts 03-08 were not re-run.
+- **No v0.3 data generation**: ProactiveAVInteraction and OCRScreenGrounding data not created.
+- **Documentation and examples only**: Added schema docs, task mapping, source provenance,
+  validation docs, model backend docs, release history, limitations, and runnable examples.
+- **Remote smoke tests**: Verification scripts for remote HF upload integrity.
+
+## Files Added/Updated (v0.2.1)
+
+- README.md (updated)
+- docs/SCHEMA.md, TASK_MAPPING.md, SOURCE_PROVENANCE.md, VALIDATION.md
+- docs/MODEL_BACKENDS.md, LIMITATIONS.md, RELEASES.md, V0_2_1_RELEASE_NOTES.md
+- examples/load_streaming_omni_dataset.py, stream_remote_dataset.py
+- examples/inspect_timeline.py, validate_one_row.py
+- scripts/15_remote_hf_smoke_test.py
+
+## Not Changed
+
+- All parquet files in data/
+- All report files from v0.2
+- Source media and intermediate files