Request access to the code-only research reference
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This repository contains code and methods documentation only. It does not grant access to training data, videos, model checkpoints, API services or credentials. Requests are automatically approved after the standard Hugging Face access form.
Log in or Sign Up to review the conditions and access this dataset content.
Video Agent SFT Pipeline — Code-Only Reference
This repository documents the engineering workflow used for video-agent SFT data preparation, real tool execution, reflection continuation, and quality review in the Thyme / ThymeV2 research workspace. It contains an allowlisted code snapshot, configurable example settings, and aggregate workload notes. It contains no training dataset or media.
The repository is gated with automatic approval. The description is visible, while downloading code requires the standard Hugging Face access request. Gating is an access mechanism, not a grant of rights to any source dataset.
What is included
| Directory | Contents |
|---|---|
code/src/video_crop_pipeline/ |
Original-video sampling, timestamped crops, clip-memory transitions, teacher rollout, blind review, step review, trajectory products, and Parquet serialization |
code/extensions/ |
Whole-trajectory text consistency review and exact exported-supervision checks |
code/reference/reflection/ |
Real prefix replay followed by live recovery, bounded retries, controller-feedback removal, and accepted-target collection |
code/reference/normal_run_reannotation.py |
Ordinary reannotation driver reference |
code/reference/dense_qa/ |
Original-QA grouping into shared-video multi-turn pure-answer conversations |
code/runtime/ |
Swift templates preserving tool supervision while supporting dense-video QA |
code/verification/ |
Native CPU encoder, loss-mask, collator and packing admission checks |
code/configs/ |
Placeholder paths and configurable example budgets; no real access configuration |
video-agent-sft-code-reference-20261012.zip contains the same code and documentation in one download. CODE_PROVENANCE.json records original and sanitized code fingerprints; PACKAGE_MANIFEST.json lists the released file fingerprints. Private workstation paths and one private endpoint/credential integration are parameterized in the released copy. No running production files were changed.
What is deliberately absent
No videos, extracted frames, image files, audio, QA examples, answers, sample-ID lists, Parquet or JSONL datasets, generated trajectories, API request/response logs, model checkpoints, tokens, private keys, certificates, or private session histories are included. File names such as sources.jsonl inside code are references to inputs that a user must provide locally, not files distributed here.
Core workflow
- Resolve an original training question, options, immutable reference answer, and verified video identity.
- Present a timestamped 64-frame overview. Obtain extra evidence through real temporal crop requests rather than invented observations.
- Preserve actual tool feedback and clip-memory state, including failures and rollback behavior.
- Require a completed final answer consistent with the original reference; assess open-answer equivalence separately when that policy is enabled.
- Run blind multimodal review, per-step supervision review, and a final text consistency gate on the actual exported targets.
- Export per-step SFT prefixes with matching visible media. Failed or replayed historical actions are not automatically positive supervision.
- Verify structure and representative native Swift encodings before a release is marked ready.
For a reference overview and setup details, see WORKFLOW.md, PROVIDER_INTEGRATION.md, and WORKLOAD.md.
Real reflection, rather than a stock sentence
The continuation adapter re-executes a recorded crop prefix and validates actual returned results, active clip indices, and rollback state. A live teacher then makes new decisions from that state. Useful recovery may shorten an oversized interval, change FPS, release an irrelevant active clip, inspect an action boundary, or revisit an unresolved entity/time relation. Acceptance requires useful newly executed evidence and a consistent completed answer. The original wrong final answer is not supplied as an answer to imitate, and old erroneous thinking is not stitched into a positive new target.
The controller's annotation-only feedback is removed from the final training context without rewriting the accepted actor targets. Whole-export review is repeated after this normalization. The reference/reflection/ scripts are a snapshot of a job-specific workflow and require local path/input/provider integration; they are not a turnkey service with embedded credentials.
Some legacy reflection helpers remain in the core snapshot because the original production module imports them. Their presence is historical, not a recommendation to concatenate a wrong chain of thought with a fixed “Wait” sentence. The genuine replay-and-recover path is the extension described above.
Ordinary and dense-video supervision
The agentic branch uses 64 overview frames plus task-dependent real crops. Each tool SFT row preserves the current visible memory and supervises the original last target under the Thyme masking convention. A middle tool target does not need a final answer of its own.
The dense-QA branch groups existing questions for the same verified original video, up to five questions per conversation. The video appears once; earlier questions and original answers become the prefix for later questions. All assistant answer letters are supervised. No tool calls or generated reasoning are added to this branch. Nominal sampling is 2 FPS with a 2,048-frame cap; longer videos retain uniform full-timeline coverage at lower effective FPS. These training inputs must not be confused with frames submitted to Codex during text-only review.
Workload and usage accounting
The project involved repeated review of roughly 10,000–15,000 SFT rows per dataset version, plus annotation/repair tasks and rollout diagnosis. One completed full review records 14,606 samples across 1,077 verified batches. A subsequent full rereview was partitioned into 14,262 rows / 617 planned batches; retained local completion receipts confirm 69 batches / 588 rows in that later pass. The full rereview input partition contains approximately 28.49 million text characters before model-specific tokenization; its preparation does not prove all of it was read. These overlapping inspections are not that many unique training examples.
Larger jobs were partitioned into 10–30 Codex task windows in the recorded workflow. Separately, external Gemini/Seed rollout jobs used up to 64 Python workers and bounded repeated attempts. Neither worker count nor review-batch count is an OpenAI billing metric. Task-level and provider-level usage must be separated; a compact final classification can still involve a long input history and multiple internal turns.
A researcher recollection of roughly 1,500 annotation/repair tasks is not a verified count of successful unique trajectories. No total model-request, token, monetary or subscription-equivalent estimate is asserted. The repository does not substantiate 256 simultaneously active Codex agents. See WORKLOAD.md for completed-versus-planned work and the scope of each figure.
Setup and limitations
Use Python 3.10 or newer and FFmpeg/ffprobe. The core package can be installed with pip install -e ./code; video-crop-pipeline --help exposes its CLI. Actual API rollouts require your own authorized provider clients, API endpoints and credentials. The original authenticated Gemini/Seed gateway implementation and certificates are not distributed. Swift encoding checks additionally require a compatible Swift/Transformers/Qwen processor installation, PyAV and qwen-vl-utils; model assets are not included.
This is a code/methods reference, not a trained model, a dataset download, a guarantee of clean annotation, or an authorization to train from any provider's outputs. Dataset rights and provider terms remain separate responsibilities. No blanket new license for third-party components is assigned; existing notices must be retained.
- Downloads last month
- 26