PaperlayerView exact revision on arXiv

Source-linked AI summary

DataClaw0: Agentic Tailoring Multimodal Data from Raw Streams

Cong Wan, Zeyu Guo, Zijian Cai, Jiangyang Li, SongLin Dong, Lin Peng, Xiangyang Luo, Zhiheng Ma, Yihong Gong

arXiv:2606.21337v2cs.LGcs.AI

TL;DR

Raw multimodal streams are difficult and costly to convert into reliable, intent-specific supervision. DataClaw0 learns this conversion and shows that joint cross-domain tailoring surpasses per-domain experts only above a capacity threshold near 18B parameters.

  • Problem

    Converting noisy raw multimodal streams into reliable, intent-specific supervision remains difficult and costly because direct annotation and repeated proprietary VLM queries are limited.

  • Method

    DataClaw0 builds supervision through deterministic factual-anchor extraction followed by VLM-based semantic synthesis, then trains intent-conditioned models with SFT and GRPO.

  • Results

    Above a capacity threshold near 18B parameters, joint training beats per-domain sharding, while below it sharding wins and joint training recovers 59% of in-domain performance on withheld domains.

  • Takeaways & Limitations

    Per-domain experts are preferable below the threshold, whereas a single joint model is preferable above it because cross-domain transfer outweighs domain competition.

  • Takeaways & Limitations

    The findings are limited to five domains and the studied schema family, and the capability is bootstrapped from a proprietary teacher.

Abstract

from arXiv · show

Raw multimodal streams are abundant but noisy, redundant, and unaligned with any particular training objective. Turning them into supervision today means either brittle heuristics or repeatedly querying a proprietary vision-language model, a cost that recurs with every new sample. We ask whether this conversion can instead be learned once and reused, and formalise intent-conditioned Data Tailoring: given a raw stream and a high-level intent, a model must return schema-aligned, evidence-grounded training instances. Training DataClaw0 at 4B, 9B and 27B, we find that whether five heterogeneous domains should share one model depends on capacity. A jointly trained model is worse than per-domain experts at the two smaller scales and better at the largest, placing the crossover near 18B parameters. Matched-data comparisons, in which the joint model sees exactly the same data per domain as that domain's expert, attribute the reversal to cross-domain transfer rather than to data volume: it gains most where a domain is data-poor, and recovers 59\% of in-domain performance on domains withheld from training entirely. Downstream post-training reproduces this ordering on GUI navigation, action video generation and spatio-temporal VQA, and the joint configuration is also the cheaper to deploy, serving one model instead of five. Github: https://github.com/vancyland/DataClaw0

1 Introduction

DataClaw0 reframes multimodal data processing as a learnable, intent-conditioned capability that converts noisy raw streams into structured supervision. Its two-stage pipeline grounds outputs in deterministic factual anchors before adding generative semantic reasoning, with performance and deployment strategy dependent on model capacity.

  • Motivation: Raw multimodal streams are abundant but noisy, redundant, weakly structured, and lack explicit supervision signals for efficient knowledge acquisition or model post-training.Distilling these continuous, high-entropy streams into structured, high-density knowledge is presented as a central multimodal data-engineering challenge.
  • Motivation: Recurring proprietary VLM annotation costs motivate treating high-quality multimodal data production as a learnable capability.Each additional sample incurs another charge, making production expense scale linearly with the amount of data required.
  • Method: DataClaw0’s two-stage construction pipeline first extracts deterministic Factual Anchors, then uses strong VLMs for long-range logical chaining over those anchors.Anchors capture object states, temporal boundaries, OCR text, and GUI interaction events, while the second stage adds multidimensional reasoning traces.
  • Framework: DataClaw0 transforms high-entropy streams into intent-aligned structured outputs and is trained at 4B, 9B, and 27B to separate architecture effects from capacity effects.The framework combines Supervised Fine-Tuning (SFT) with Group Relative Policy Optimization (GRPO).
  • Results: 27B surpasses substantially larger proprietary annotators on all three DataClaw0-val axes, whereas the 4B and 9B variants do not.DataClaw0-val includes an Intent subset for vague-intent concretization and evaluates structured data refinement.
  • Results: Whether five domains should share one model depends on capacity: across 4B, 9B, and 27B backbones, per-domain experts are better at the smaller scales.The introduction frames this comparison through Omni and Expert deployment paradigms.

2 Related Work

Related work spans multimodal large language models, which extend language-model capabilities to visual perception, grounding, and multimodal reasoning, and agentic AI, which enables goal-driven planning, tool use, environmental interaction, and iterative refinement.

  • Multimodal Large Language Models: Multimodal large language models combine visual encoders, multimodal projectors, and language backbones to support visual perception, grounding, and multimodal reasoning.They extend the reasoning, instruction-following, and generalization capabilities demonstrated by text-only language models.
  • Agentic AI: Agentic AI extends language models into goal-driven systems that plan, use tools, interact with environments, execute multi-step workflows, and refine results through feedback.OpenAgents, LangChain, AutoGPT, and OpenClaw exemplify systems and frameworks for orchestrating tools.

3 Method

DataClaw0 learns intent-conditioned multimodal data tailoring: it filters and reorganizes noisy raw streams into schema-aligned, evidence-grounded assets. Its pipeline combines factual-anchor extraction, VLM synthesis, supervised training, and rule-driven GRPO optimization.

  • Intent-conditioned tailoring: DataClaw0 converts lengthy raw multimodal streams and high-level intents into customized outputs that are semantically accurate and strictly schema-conforming.The agent must filter redundant background noise and reorganize information into predefined formats such as JSON, Markdown, or action-code logic.
  • Training-data construction: The two-stage construction pipeline extracts timestamped factual anchors with lightweight experts, then uses a strong VLM to synthesize intent-conditioned structured supervision.Anchors provide grounding signals that reduce hallucinations in long-sequence annotation.
  • Training-data construction: Gemini-3.1-Pro is queried once per training example to build the SFT corpus, while per-domain lightweight deterministic components perform anchor extraction.The synthesis engine is never invoked during tailoring; the anchor extractor is an ensemble rather than a single model.
  • Optimization: After SFT, rule-driven GRPO uses deterministic formatting, anchor-alignment, and verbosity rewards to improve output validity, grounding, and concision.The policy uses group-normalized advantages, a clipped objective, and a KL regularizer against the reference model.
  • Deployment and evaluation: DataClaw0 supports unified omni deployment for cross-domain flexibility and domain-decoupled experts for specialization, with inference including schema-constrained policy inference and grounding verification.Evaluation includes JSON validity, schema-field correctness, textual semantic alignment, and trajectory-shape similarity.

4 Experiments

Experiments show that joint training overtakes domain sharding only above a capacity threshold, because cross-domain transfer offsets data scarcity. Matched-data, held-out-domain, and downstream evaluations support transfer, generalisation, and practical utility.

  • Capacity threshold: 14.6 points of Overall separate sharding from joint training at 4B, 11.0 at 9B, while joint training leads by 5.9 at 27B, with crossover near 18B parameters.The ordering reverses across backbone sizes, making the crossover a capacity threshold rather than a property of one model.
  • Capacity threshold: At 27B, joint training gains most on AIGC and Education, while specialisation remains slightly better on data-rich Daily Life.Semantic scores are 80.36 vs. 71.83 for AIGC, 82.51 vs. 74.96 for Education, and 61.24 vs. 61.82 for Daily Life.
  • Downstream validation: 43.1 vs. 39.5 on GUI Step Success Rate and 56.7 vs. 53.4 on VQA Partial Accuracy show DataClaw0-27B-O leads on downstream metrics.End-to-end gaps are TSR 19.8 vs. 14.2 and VQA Overall 37.6 vs. 31.5; benchmark Overall correlates with downstream gain at Pearson r = 0.97 and Spearman ρ = 0.98.
  • Cross-domain transfer: The matched-data joint model is ahead on four of five domains, showing that its advantage comes from other domains rather than additional in-domain data.Daily Life is the sole exception.
  • Cross-domain transfer: 59% of the gap is recovered on average when each domain is withheld entirely, including 69% on Daily and 67% on Education but 50% on GUI.The joint model recovers this performance without seeing any examples from the held-out domain.

5 Conclusion · A DataClaw0 Dataset and Benchmark · A.1 Dataset Overview

DataClaw0 converts raw multimodal streams into intent-conditioned, schema-aligned, evidence-grounded training data and evaluates this capability across five diverse domains. Its joint-model advantage emerges above a capacity threshold near 18B parameters, while the dataset and benchmark emphasize transparent construction, reproducibility, and leakage control.

  • 5 Conclusion: Above 18B parameters, one joint model outperforms per-domain sharding, while below that threshold the five domains compete and experts win.The joint model recovers 59% of in-domain performance on domains withheld from training entirely.
  • 5 Conclusion: DataClaw0 amortises a proprietary teacher’s capability into a smaller open-weight model rather than creating capability the teacher lacks.The authors do not claim generalisation beyond the five studied domains and schema family.
  • A DataClaw0 Dataset and Benchmark: The appendix documents the corpus, construction pipeline, DataClaw0-val benchmark, fuzzy-intent subset, and structured metrics to support transparent, reproducible evaluation.These materials cover both data construction and evaluation protocols.
  • A.1 Dataset Overview: The corpus spans five domains—Daily Life, Education, Embodied AI, World Models/AIGC, and GUI Agents—with raw streams, factual anchors, and verified structured outputs.The streams range from long GUI operation logs to physical manipulation trajectories and instructional videos.
  • A.1 Dataset Overview: The domains represent distinct multimodal entropy, including GUI text and actions, embodied states and trajectories, procedural videos, and scene-layout or motion structures.This diversity covers different forms of multimodal information and temporal structure.
  • A.1 Dataset Overview: Domain-specific intents define downstream schemas for GUI navigation, spatio-temporal VQA, embodied action trajectories, and video-generation supervision.Each domain contains a set of tailoring intents specifying the target use case and output schema.
  • A.1 Dataset Overview: Validation samples are source-disjoint from training whenever identifiable stream or task IDs exist, with repeated templates and near-identical action sequences removed for GUI and embodied data.Raw videos, GUI sessions, and embodied trajectories in DataClaw0-val are excluded from SFT and GRPO pools.

A.2 A Complete Data Construction Case · A.3 Data Construction Pipeline

DataClaw0 converts raw multimodal streams into intent-conditioned, schema-aligned supervision by extracting factual anchors and synthesizing structured training instances. Its pipeline separates deterministic preprocessing from constrained multimodal synthesis, validation, and corpus export.

  • A.2 A Complete Data Construction Case: DataClaw0 transforms high-entropy multimodal observations into structured, task-specific instances rather than merely captioning videos or screenshots.The construction principle extracts factual anchors, conditions generation on user intent, and produces schema-compliant outputs suitable for downstream training.
  • A.2 A Complete Data Construction Case: In embodied manipulation, the model infers the next high-level subtask from temporal evidence and object-state transitions.After the robot places yellow beads into a blue bowl, the remaining green block becomes the predicted next object to pick up.
  • A.2 A Complete Data Construction Case: For long-horizon GUI composition, DataClaw0 aligns compatible fragments, orders them temporally, and rejects distractors from unrelated applications.It identifies Fragments B, A, and C as a coherent Excel task: formatting precedes entry, and typo correction follows erroneous header creation.
  • A.2 A Complete Data Construction Case: For daily-life environmental understanding, the model generates grounded reasoning questions about scene layout, visible objects, and environmental cues.The example infers a living area or study space from a wooden table, books, a glass-like object, and other household items.
  • A.3 Data Construction Pipeline: The construction pipeline first discovers candidate segments and packages sampled frames, actions, descriptions, timestamps, or frame ranges into compact requests.These steps convert raw videos or GUI trajectories into compact, verifiable construction units before semantic synthesis.
  • A.3 Data Construction Pipeline: It then performs intent-conditioned VLM synthesis under task-specific prompts and strict schemas, validates generated JSON and temporal references, and materializes media.The final stage exports normalized training samples alongside artifact mapping files for downstream training and inspection.

A.3.1 Robot Data Pipeline … A.4.2 Reference Output Construction

DataClaw0 converts raw robot and GUI streams into schema-aligned, evidence-grounded training samples through deterministic preprocessing and model-based semantic synthesis. DataClaw0-val evaluates this tailoring ability across five multimodal scenarios using structured reference outputs aligned with both evidence and construction intent.

  • A.3.1 Robot Data Pipeline: DataClaw0 transforms robot videos into fault-recovery and operation-understanding data by selecting informative windows, eliciting structured supervision, and materializing referenced media.The fault branch creates counterfactual stalled inputs grounded in successful continuations, while the understanding branch produces temporally grounded reasoning examples.
  • A.3.2 GUI Multi-event Construction Pipeline: The GUI pipeline filters likely multi-event trajectories, decomposes them into subtask fragments, and shuffles fragments for long-horizon task reconstruction.It targets fragment ordering, cross-fragment state tracking, and task-level composition rather than simple action segmentation.
  • A.3.3 Unified Output Format and Artifact Management: Robot and GUI outputs store semantic supervision in lightweight JSONL while materializing large media assets separately through artifact metadata.This separation supports relocating or regenerating media without changing logical sample content.
  • A.3.4 Validation and Quality Control: DataClaw0 combines deterministic preprocessing with VLM synthesis, then validates schemas, temporal ranges, media artifacts, and GUI fragment structure.The pipeline also supports resumable processing and intermediate JSONL records for auditability and reproducibility.
  • A.4 DataClaw0-val: DataClaw0-val measures whether models can transform high-entropy multimodal inputs and explicit tailoring intents into structured training-data instances.Each example specifies multimodal input, tailoring intent, target schema, and reference output.
  • A.4.1 Benchmark Composition: DataClaw0-val contains 200 examples across AIGC/World, Daily Life, Education, Embodied AI, and GUI Agents, while DataClaw0-Intent contains 120 underspecified-intent examples.The categories cover generative-video segment selection, navigation, teaching-material reconstruction, robot-operation data construction, and GUI task reconstruction.
  • A.4.2 Reference Output Construction: Reference outputs use task-specific JSON schemas and only indices or fragment labels present in the supplied inputs.Video and image indices are zero-based over sampled frames, while GUI labels come from candidate fragments.
  • A.4.2 Reference Output Construction: References are constructed to satisfy schema validity, evidence grounding, and stated data-construction intent, with validation data separated from the training corpus.Outputs must perform the requested tailoring task rather than merely caption or summarize the media.

A.4.3 DataClaw0-Intent: Fuzzy-intent Stress Test … B.2 Inference Architecture & Deployment Paradigms

DataClaw0-Intent tests whether models can infer useful structured training objectives from vague multimodal instructions, while the broader system combines hierarchical evaluation, rule-driven GRPO, and flexible inference for valid, grounded outputs.

  • A.4.3 DataClaw0-Intent: Fuzzy-intent Stress Test: Fuzzy-intent evaluation requires inferring a useful data-construction objective from raw multimodal input and producing structured output despite underspecified instructions.Unlike standard schema following, the model must determine the training-data type, informative evidence, and output fields.
  • A.4.3 DataClaw0-Intent: Fuzzy-intent Stress Test: Because fuzzy-intent videos admit multiple useful training instances, DataClaw0-Intent uses a user study rather than exact-match scoring against one reference.Possible conversions include future-frame prediction, action anticipation, event segmentation, and visual reasoning.
  • A.4.3 DataClaw0-Intent: Fuzzy-intent Stress Test: 120 fuzzy-intent samples were each judged by five raters across four five-point Likert dimensions, producing 600 judgements from 100 raters.Submissions failing two attention checks were discarded, and Krippendorff’s α was 0.71.
  • A.4.4 Metrics: The hierarchical multimodal-data score uses a hard JSON-validity gate and weights λ1 = 0.20, λ2 = 0.35, and λ3 = 0.45.Additional components measure schema integrity, semantic similarity, and spatio-temporal trajectory-shape similarity.
  • B.1 Rule-Driven Reinforcement Learning via GRPO: GRPO avoids a comparable-size Critic model and uses deterministic rule-based rewards tailored to strictly structured outputs.Rewards verify schema compliance, measure spatio-temporal anchors, and regulate reasoning length.
  • B.1 Rule-Driven Reinforcement Learning via GRPO: K = 50 alignment points and τs = 0.10 support the spatio-temporal anchor reward’s trajectory-shape comparison.Trajectories are normalized, resampled, compared using MAE, and converted into a similarity-based reward.
  • B.1 Rule-Driven Reinforcement Learning via GRPO: SFT initialization followed by rule-reward-driven GRPO stimulates agentic structured-data tailoring, particularly spatio-temporal grounding through Ranchor.GRPO samples diverse candidate outputs, scores them jointly, normalizes rewards within each group, and updates the policy using relative advantages.
  • B.2 Inference Architecture & Deployment Paradigms: The inference architecture converts raw multimodal streams and user intents into schema-aligned structured outputs while emphasizing validity, grounding reliability, and deployment flexibility.It is designed to transform the GRPO-aligned policy into a practical structured tailoring agent.

B.2.1 Core System Architecture · B.2.2 Omni vs. Expert Deployment Paradigms · C Downstream Evaluation

DataClaw0 uses three decoupled modules to transform raw multimodal streams and intents into structured, schema-aligned, factually grounded outputs. It supports both unified omni deployment and modular domain-specific experts, but the supplied passages do not detail downstream evaluation results.

  • B.2.1 Core System Architecture: Three decoupled modules comprise DataClaw0’s inference process.The modules are multimodal ingestion and intent parsing, schema-constrained inference, and structural verification and grounding.
  • B.2.1 Core System Architecture: The ingestion module converts raw multimodal streams Xraw and user intents I into structured context sequences for the policy model.Rule-based scripts currently segment long video trajectories into semantically coherent clips using raw dataset annotations.
  • B.2.1 Core System Architecture: Future ingestion will use long-video understanding to adaptively identify event boundaries and intent-relevant sub-sequences.This extends the current scalable preprocessing pipeline beyond rule-based segmentation.
  • B.2.1 Core System Architecture: Schema-aware decoding suppresses invalid continuations that violate the target schema Φ.The schema-constrained inference engine deploys policy model πθ to generate structured outputs.
  • B.2.1 Core System Architecture: Structural verification cross-checks generated UI components, coordinates, timestamps, and action trajectories against factual anchors A.Quantitative evaluation metrics support post-hoc structural and semantic verification, ensuring semantic grounding and structural reliability.
  • B.2.2 Omni vs. Expert Deployment Paradigms: DataClaw0-O jointly trains one model across all domains to serve diverse cross-domain intents I.It is positioned as a generalist agent within a single model.
  • B.2.2 Omni vs. Expert Deployment Paradigms: DataClaw0-E selects domain-specific tailoring agents according to the target scenario, domain setting, or user configuration.This architecture improves domain-specific robustness and permits modular updates.

C.1 Detailed Analysis of Downstream Application … D Additional Experiments and Ablations

DataClaw0 produces compact, schema-aligned supervision that improves downstream multimodal adaptation across GUI navigation, action video generation, and spatio-temporal VQA. Under matched data budgets and filtering, it generally transfers better to complete task-solving and offers a practical alternative to proprietary annotation.

  • C.1 Detailed Analysis of Downstream Application: Downstream evaluation covers long-horizon GUI navigation, action video generation, and spatio-temporal VQA, spanning action planning, visual dynamics generation, and temporal understanding.These tasks test different output spaces and reasoning requirements.
  • C.1 Detailed Analysis of Downstream Application: Self-processed data yields limited improvements over zero-shot base models, whereas Gemini-processed and DataClaw0-processed data produce substantial downstream gains.The comparison indicates that high-quality structured supervision is important for downstream adaptation.
  • C.1 Detailed Analysis of Downstream Application: Gemini performs slightly better on some partial or step-level metrics, while DataClaw0 achieves higher GUI TSR and VQA Overall Accuracy.The results suggest richer intermediate supervision favors local correctness, whereas DataClaw0 better supports complete task-solving.
  • C.1 Detailed Analysis of Downstream Application: In action video generation, DataClaw0 achieves lower FVD and higher Contact mAP, indicating improved visual quality and physical-interaction modeling.These metrics complement the GUI and VQA comparisons.
  • C.2 Unified Refinement Protocol: The unified refinement protocol processes identical raw streams with multiple annotators, applies the same schema and filters, samples equal valid instances, and fine-tunes the same downstream model.The protocol is designed to compare data quality rather than data quantity.
  • C.3 GUI Navigation: GUI navigation fine-tunes Qwen3.5-4B to predict actions from screenshots, task instructions, and interaction history, reporting Step Success Rate and Task Success Rate.SSR measures step-level correctness; TSR measures successful completion of the full task within a step limit.
  • C.4 Action Video Generation: Action video generation fine-tunes Wan2.2-I2V-5B to generate short videos from an input image and text prompt, with outputs compared across refined datasets.The section also provides the fine-tuning setup and visualization comparisons.
  • C.5 Spatio-temporal VQA: Spatio-temporal VQA fine-tunes Qwen3.5-4B to answer questions from video frames, evaluating both Partial Accuracy and Overall Accuracy.Partial Accuracy allows partially correct structured answers, while Overall Accuracy requires the complete answer to match.

D.1 Scaling Laws and Emergent Diversity

DataClaw0’s unified architecture is capacity-limited: expert routing wins at 4B and 9B, but the unified model surpasses experts at 27B. DataClaw0 also produces broader semantic coverage and performs strongly on vague, high-level instructions.

  • Scaling Laws: At 9B, DataClaw0-O fluctuates from 53.60 to 47.23, 57.84, and later oscillations, while DataClaw0-E scales monotonically to 68.86.The instability is attributed to gradient conflict from forcing shared weights across five heterogeneous extraction-task distributions; domain experts avoid this conflict.
  • Scaling Laws: At 27B, the unified model leads the Expert ensemble by 5.9 points of Overall, reversing the 14.6-point and 10.5-point Expert advantages at 4B and 9B.The 27B model also follows a monotone scaling curve, unlike the oscillating 9B unified model.
  • Emergent Diversity and Intent Comprehension: DataClaw0-refined data covers semantic space substantially more broadly and evenly than Raw Data or Qwen3.5-9B refinements, revealing new clusters and long-tail patterns.The comparison uses semantic embeddings and t-SNE visualizations across the three data sources.
  • Emergent Diversity and Intent Comprehension: On the DataClaw0-Intent subset of vague, high-level instructions, DataClaw0 significantly surpasses Qwen3.5-9B and approaches Gemini 3.1-pro-preview.The benchmark is designed to quantify intent comprehension under fuzzy instructions.

E Qualitative Analysis · E.1 Successful Case Studies · E.2 Failure Cases

DataClaw0’s qualitative analysis illustrates its end-to-end tailoring process across representative domains while exposing failures in temporal grounding. In particular, outputs can be semantically correct yet narratively reverse the true frame chronology because of temporal hallucination.

  • E.1 Successful Case Studies: DataClaw0’s successful case studies show the end-to-end process from raw input and user intent to tailored output across representative domains.Each case presents the raw input, user intent, and DataClaw0 output.
  • E.2 Failure Cases: DataClaw0 can construct a semantically correct navigation sample whose question asks how to move from the workspace to the bed and whose answer describes that transition.The failure concerns the reasoning narrative rather than the constructed sample’s semantic content.
  • E.2 Failure Cases: The CoT may reverse the actual frame order, describing movement from workspace to bed when frames 0 →20 show the opposite temporal direction.
  • E.2 Failure Cases: Temporal hallucination occurs when multimodal models infer a coherent event flow from spatial cues instead of the exact chronological order of frames.The passage attributes this behavior largely to the base model’s reasoning prior and identifies strict temporal grounding as intrinsically difficult.
  • E.2 Failure Cases: Figure 10 compares t-SNE embeddings of tailored outputs, the raw input distribution, and the untrained backbone using the 9B model.The figure is noted as requiring regeneration for the 27B configuration.
  • E.2 Failure Cases: The qualitative failure-analysis prompt requests one Observe Surroundings Future Generation example from a gameplay video with Chain-of-Thought reasoning.
  • E.1 Successful Case Studies: A successful world-model or video-generation tailoring case is highlighted among the qualitative examples.

F Amortised Cost Analysis

The analysis compares self-hosted DataClaw0-9B/27B with per-sample API annotation through an upfront-cost versus marginal-cost break-even model. Self-hosting becomes cheaper for repeated production under favorable accelerator pricing, while the joint deployment configuration reduces serving costs and preserves on-premise, provider-independent operation.

  • Cost model: Self-hosting pays an upfront distillation and training cost, then produces samples at marginal cost clocal, versus n capi for API annotation.The break-even comparison is defined by Cfixed + n clocal = n capi, where Cfixed includes teacher calls and training compute.
  • Backbone economics: At 27B, training costs $3,740 and marginal serving costs $0.027 per sample, compared with $2,360 and $0.009 at 9B.The larger backbone is roughly three times more expensive to train and produces roughly one third as many samples per accelerator-hour when considered alone.
  • Deployment cost: Deployment favors the joint configuration because one 27B model replaces five resident 9B domain experts and requires no router.Thus the relevant deployment comparison is one 27B model against five 9B models, rather than one model against one model.
  • Break-even: Against Gemini-3.1-Pro at $0.038 per sample, 9B breaks even at roughly 80K samples and 27B at roughly 324K.Against Kimi-K2.6, 9B breaks even at 327K samples, while 27B never breaks even because its marginal cost exceeds the API price.
  • Sensitivity and scope: The 9B investment recovers between 54K and 149K samples under typical accelerator pricing, whereas 27B pays off only with competitively priced accelerators and list-tier API billing.The analysis supports a narrower claim: self-hosting is cheaper for repeated production on competitively priced accelerators, not universally.
  • Non-monetary considerations: Self-hosted tailoring keeps sensitive raw streams on-premise and removes rate limits and provider dependence during continuous large-scale production.These benefits are not captured by the monetary break-even equation and matter for GUI logs and robot trajectories containing potentially sensitive information.

G Limitations and Future Work

The study’s claims are limited by dependence on a proprietary teacher, task-aligned evaluation, single-reference sequence matching, and a narrow empirical scope. Future work targets broader validation, improved reproducibility and provenance practices, and autonomous supervision construction through environment interaction.

  • Teacher dependence: DataClaw0 distills Gemini-3.1-Pro, so it does not claim capabilities absent from the proprietary teacher; its practical contribution is amortising them into a 9B open-weight model.The resulting marginal cost per sample is local inference, while cross-domain transfer enlarges what a fixed teacher budget buys.
  • Evaluation validity: DataClaw0-val shares the construction pipeline’s schema, anchor definitions, and intent taxonomy, so its score measures task compatibility rather than neutral data-refinement ability.The authors have not established that DataClaw0-val predicts downstream gain and therefore treat downstream post-training as the primary criterion.
  • Metric limitations: Single-reference sequence matching can score valid alternative action orderings as errors, making reported sequence numbers a lower bound on tailoring quality.The proposed remedies are a goal-conditioned criterion or reference sets admitting multiple valid orderings.
  • Empirical scope: Results across 4B, 9B, and 27B backbones locate the capacity threshold between 9B and 27B but cannot characterize its shape or generality across model families.A denser sweep or models beyond 27B could test whether the joint-training advantage grows or saturates; varying only dense backbones leaves parameter count, pretraining data, and mixture-of-experts effects confounded.
  • Reproducibility and provenance: Unreleased implementation details and potentially identifying content in egocentric video and GUI logs limit reproducibility and raise data-provenance concerns.The authors intend to release the construction pipeline and evaluation harness, while using sources released for research use.
  • Future work: A natural extension is autonomous construction, with the model exploring embodied, game, or GUI simulations and creating supervision from its own interactions.This would move DataClaw0 beyond refining, filtering, and annotating user-provided raw streams toward interaction-driven data construction.
Loading 2606.21337v2…