Pith. sign in

REVIEW 2 major objections 3 minor 45 references

One protocol makes live model training steerable and auditable.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 20:59 UTC pith:GMG4X3TJ

load-bearing objection A reusable control plane for live training that's honestly scoped and well-documented; fix the flush/durability wording and the abstract's 'safe' claim and it's solid. the 2 major comments →

arxiv 2607.18314 v1 pith:GMG4X3TJ submitted 2026-07-17 cs.LG

Interactive Training 2: Auditable Control Plane for Live Model Training

Wentao Zhang , Xuanhe Pan , Han Zhou , Yang Lu , Yuntian Deng This is my paper
classification cs.LG
keywords live trainingcontrol planeauditable trainingLLM-guided trainingexperiment trackinginteractive trainingreinforcement learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper seeks to establish that steering a live training run—changing a learning rate, rebalancing data, requesting a checkpoint—can be reduced to a reusable, auditable protocol rather than per-experiment custom code. It presents a control plane in which training applications declare typed settings and structured actions, any controller (human, script, heuristic, or LLM agent) submits requests through the same queue, and the training loop itself chooses the safe points where requests are applied and records every outcome. If the approach holds, researchers would no longer rebuild monitoring, control, and logging machinery for each workflow; instead they register knobs and call a session at safe boundaries, and all interventions become inspectable in a unified chronological journal. The paper demonstrates this across five NLP and reinforcement-learning workflows and releases the code and traces as reusable evidence.

Core claim

The central claim is that a single control-plane abstraction—typed knobs and actions registered by the training application, a request queue drained at application-chosen control points, and an ordered event journal connecting plans, requests, outcomes, metrics, and checkpoints—can represent many kinds of live changes and give every controller a uniform interface. The training loop retains ownership of every model update; a long-lived session object validates and clamps requested values, applies the change at the next control point before any further training step, and returns a control object that the integration translates into framework-specific semantics. Every request carries a source,

What carries the argument

The central object is the TrainingSession: a long-lived component that owns a request queue, registered knobs (typed getter/setter pairs with optional range and step constraints), structured actions, and an ordered event journal. Its step(metrics) call is the safe control point: it records new metrics, runs attached controllers on schedule, drains the queue one request at a time, invokes the matching handler with clamping, records a result for each request, and returns a StepControl that tells the loop whether to stop, evaluate, save or load a checkpoint, reset a module, or update a setting. This machinery guarantees that when the call returns, every accepted change is in place before the ne

Load-bearing premise

The system's safe operation rests on application developers choosing semantically safe control points and writing correct setter and action handlers; the session only checks types and ranges and records results, so a poorly placed control point or incorrect setter will journal a success while actually damaging the run.

What would settle it

Take a released trace and replay it with a deliberately adversarial controller that requests extreme in-range settings at every control point; if the journal records success for every request while the run silently diverges or the setting never takes effect (e.g., inspecting the optimizer's hyperparameters immediately after the request shows no change), then the guarantee that 'every accepted change is already in place' is empty. A simpler check: instrument the optimizer state right after a set_knob request for a learning rate and verify the actual optimizer value changed before the next step.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Different training frameworks—callbacks, optimizer wrappers, direct loops, and reinforcement-learning loops—can be steered through the same protocol without rebuilding the monitoring, control, and logging stack for each one.
  • Every intervention, whether from a human or an automated controller, leaves an auditable record: source, request contents, success or failure, and surrounding metrics and checkpoints, so experiments can be reconstructed later.
  • The cross-round journal lets an LLM agent carry reflections from one fresh round into the next plan, so failures and stalled rounds visibly influence subsequent configurations.
  • The protocol is controller-agnostic, meaning existing hyperparameter-search and dynamic-configuration algorithms can act as controllers on top of the control plane without modifying training code.
  • The released journals and traces provide a shared foundation for studying how humans and agents supervise learning in progress, including cases where later rounds recover after regressions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural extension is to add formal preconditions and postconditions to knob declarations, so the control point can check not just type and range but also semantic safety (e.g., 'the new learning rate must be positive and the optimizer must already exist'), turning the journal into a richer certificate of correctness.
  • The round-structured traces could be replayed as a benchmark for steering policies: fixing the initial conditions and journal, one could compare how often different controller policies improve the best score per action taken or token spent, something the paper does not itself quantify.
  • Because the protocol decouples controllers from the training loop, it suggests a clean way to implement multi-controller arbitration—prioritization, vetoes, approval gates, resource budgets—without touching training code, an extension the paper notes is needed for high-stakes use.
  • A measurable open question the paper leaves implicit is overhead: how much training throughput is lost per control-point call, which would matter for deciding how frequently to invoke the session in very large runs.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 3 minor

Summary. The paper presents Interactive Training 2, an open-source control plane that allows humans, scripts, heuristics, and LLM agents to steer a live training run through a shared protocol. Training applications register typed settings (knobs) and structured actions; the training loop calls a TrainingSession at application-chosen control points, where queued requests are validated, applied, and recorded. A customized Aim workspace provides live metrics, controls, and an ordered event journal. The authors demonstrate the system on five NLP and RL workflows, and release code, JSONL journals, a SHA-256 manifest, and an interactive sandbox. Score improvements are explicitly presented as context, not as controlled comparisons. The paper’s central system claim is that training can be made steerable and auditable through one reusable protocol.

Significance. If the auditability guarantee is made precise, this is a useful systems contribution: it decouples controllers from trainer-specific code, gives human and automated controllers one interface, and provides a cross-round journal connecting decisions to outcomes. The paper’s strengths include a concrete implementation, a smoke test, five released workflows, JSONL journals with a SHA-256 manifest, and a candid Limitations section that acknowledges the safety boundary and the fact that scores are not controlled comparisons. The safety limitation is explicit, but the durability of the audit trail is not; resolving that gap is necessary before the central claim can be accepted as stated.

major comments (2)
  1. [§4.2, §4.5, Appendix B] The central auditability claim is not yet supported by a durability specification. §4.2 step 4 says the loop “records a result for every request,” and §3 says the session “validates it, and records the result before continuing,” but §4.5 states that the Aim transport records events “on a background thread so writing metrics does not pause training.” No statement says an action_result is durably flushed before session.step returns, and no crash behavior is specified. If the transport is an in-memory queue or an unsynchronized background write, a crash or write failure can drop exactly the event needed to reconstruct a change. The released JSONL journals (Appendix B) are round-level summaries, so they do not restore the per-request event stream. Please specify the persistence contract for action results (e.g., synchronous durable write before the step returns, or an explicit at-least-once/
  2. [§5.2 and Appendix B] The validation claim “the request-to-result path can be checked against the released traces” is stronger than the released JSONL supports. Appendix B describes the cross-round JSONL journal as containing the initial configuration, best score and step, successful actions, and reflection—not the per-request event stream with individual action_result entries. The per-request path is visible in the Aim workspace and the Figure 1 capture, but the paper should clearly state which released artifact contains the full event stream, and how a reviewer can replay the request-to-result path from the released files alone. This is a reproducibility issue for the auditability claim.
minor comments (3)
  1. [Abstract and §3] Consider replacing “safe control points” / “safely applied” with “application-selected control points” or adding an explicit qualifier that safety is developer-provided. The Limitations paragraph already states this, but the abstract and contribution wording currently overpromise relative to the system's own guarantees.
  2. [§4.5, line “Aim keeps a permanent copy” (Appendix B)] The statement that Aim keeps a permanent copy is unsupported by any durability mechanism described in the paper. If the transport is asynchronous, “permanent” needs qualification or a reference to the durability mechanism.
  3. [Table 1 and §5] The score columns are appropriately labeled as context, and the paper should be commended for not presenting them as controlled comparisons. Consider adding a sentence in §5 reminding readers that all runs use fresh models/optimizers per round, so score trajectories reflect both the controller and the new initialization.

Circularity Check

0 steps flagged

No significant circularity: the paper presents a software protocol and records outcomes as context; no fitted parameter or self-citation is repackaged as a prediction.

full rationale

The paper's central claims are about a reusable control-plane protocol: training applications declare typed settings and actions, requests are queued, applied at developer-chosen control points, and recorded in a journal. There is no mathematical derivation whose output reduces to an input, and no fitted parameter is later reported as a prediction. The multi-round scores are explicitly framed as context rather than controlled optimization comparisons ('Scores provide context for the recorded decisions; they are not controlled comparisons of optimization algorithms'), so the empirical tables are not used to claim a derived performance advantage. The only self-citation is to the authors' earlier Interactive Training v1, and it is used to describe the predecessor system and to delineate what is new, not to justify the current protocol's correctness or uniqueness. The limitations section candidly states that application developers remain responsible for custom-action safety and that the system cannot guarantee safe or optimal actions, which confirms that the 'safe control point' premise is an acknowledged design assumption rather than a concealed circular step. The skeptical concern about Aim's background-thread event transport lacking a flush/durability guarantee is a real engineering risk, but it is not circularity: it questions whether the auditable-record guarantee holds under crash, not whether the output is equivalent to the input by construction. Accordingly, no circular step is identified and the circularity score is 0.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

No free parameters are fitted and no mathematical derivation is attempted. The central artifact rests on three domain assumptions: developers can choose safe control points; loop integrations correctly map StepControl decisions; and the background event transport preserves audit records. These assumptions are practical rather than mathematical, and the released traces give moderate support but no proof.

axioms (3)
  • domain assumption Application developers can identify semantically safe control points and implement correct setters/actions for their training loops.
    Section 3: 'The application chooses the control points, validates task-specific actions...'; Limitations: 'Applications remain responsible for checking custom actions.' If a developer chooses an unsafe point, the system records an applied result but cannot guarantee safe training.
  • domain assumption Training-loop integrations map StepControl decisions to framework semantics without altering model-update behavior.
    Section 4.2: 'The control point guarantees that when the call returns, every accepted change is already in place before the next step' assumes the integration correctly translates stop/evaluate/checkpoint/setting updates into local framework semantics. This is validated in five workflows, not proven generally.
  • domain assumption The background-thread Aim transport preserves the event records needed for the audit trail.
    Section 4.5: 'An Aim transport records events on a background thread so writing metrics does not pause training.' The paper does not state crash or durability guarantees, yet reconstructing who requested what depends on event persistence.

pith-pipeline@v1.3.0-alltime-deepseek · 9112 in / 13426 out tokens · 123717 ms · 2026-08-01T20:59:41.427763+00:00 · methodology

0 comments
read the original abstract

Experiment trackers show how training is progressing, but changing a live run still usually requires trainer-specific code. We present Interactive Training 2, an open-source control plane for steering training through a shared protocol. Training applications declare which settings and actions they expose, humans and automated controllers submit requests through the same interface, and the training loop validates and applies them at safe control points. A customized Aim workspace combines live metrics and controls with a chronological record of requests and outcomes. We demonstrate the system across five NLP and reinforcement-learning workflows. The released code and traces provide a reusable foundation for auditable human- and agent-guided training.

Figures

Figures reproduced from arXiv: 2607.18314 by Han Zhou, Wentao Zhang, Xuanhe Pan, Yang Lu, Yuntian Deng.

Figure 1
Figure 1. Figure 1: One request from plan to result in the Aim Live workspace. Enlarged crops from one illustrative BERT run show (1) the Round 2 journal and proposed change, (2) Controls and best-so-far round scores, (3) Validation loss after the setting change, and (4) Logging area showing the request, result, and saved checkpoint. ling, and reviewing training. A customized Aim workspace combines live metrics and controls w… view at source ↗
Figure 2
Figure 2. Figure 2: The shared control protocol. Training code declares settings and actions and calls TrainingSession at safe control points. Human users, scripts, and the optional LLM agent submit the same typed actions; the session validates, applies, and records them before the next training step, while Aim presents live metrics and the event journal. The lower panel shows what carries across rounds; each new round create… view at source ↗
Figure 3
Figure 3. Figure 3: Scores recorded during five example runs. Each point is the best score recorded in one fresh round. The solid line tracks the best score seen so far, the diamond and dashed line mark the Round 0 reference without the LLM agent, and filled circles mark a new best. Y-axes use task-specific raw units and fixed limits. All panels share the Round 0–10 x-scale, with Countdown ending at Round 7. Sentiment and Cou… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 2 canonical work pages

  1. [1]

    Steven Adriaensen, Andr \'e Biedenkapp, Gresa Shala, Noor Awad, Theresa Eimer, Marius Lindauer, and Frank Hutter. 2022. https://doi.org/10.1613/jair.1.13922 Automated dynamic algorithm configuration . Journal of Artificial Intelligence Research, 75:1633--1699

    doi:10.1613/jair.1.13922 2022

  2. [2]

    Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. 2019. https://doi.org/10.1145/3292500.3330701 Optuna: A next-generation hyperparameter optimization framework . In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 2623--2631

    arXiv 2019

  3. [4]

    Francesco Barbieri, Jose Camacho-Collados, Luis Espinosa-Anke, and Leonardo Neves. 2020. TweetEval:Unified Benchmark and Comparative Evaluation for Tweet Classification . In Proceedings of Findings of EMNLP

    2020

  4. [5]

    James Bergstra and Yoshua Bengio. 2012. http://jmlr.org/papers/v13/bergstra12a.html Random search for hyper-parameter optimization . Journal of Machine Learning Research, 13(10):281--305

    2012

  5. [6]

    Lukas Biewald. 2020. https://www.wandb.com/ Experiment tracking with weights and biases . Software available from wandb.com

    2020

  6. [8]

    Czarnecki, Jeff Donahue, Ali Razavi, Oriol Vinyals, Tim Green, Iain Dunning, Karen Simonyan, Chrisantha Fernando, and Koray Kavukcuoglu

    Max Jaderberg, Valentin Dalibard, Simon Osindero, Wojciech M. Czarnecki, Jeff Donahue, Ali Razavi, Oriol Vinyals, Tim Green, Iain Dunning, Karen Simonyan, Chrisantha Fernando, and Koray Kavukcuoglu. 2017. https://arxiv.org/abs/1711.09846 Population based training of neural networks . Preprint, arXiv:1711.09846

    Pith/arXiv arXiv 2017

  7. [9]

    Keller Jordan, Yuchen Jin, Vlado Boza, Jiacheng You, Franz Cesista, Laker Newhouse, and Jeremy Bernstein. 2024. https://kellerjordan.github.io/posts/muon/ Muon: An optimizer for hidden layers in neural networks

    2024

  8. [10]

    Ilya Loshchilov and Frank Hutter. 2019. https://openreview.net/forum?id=Bkg6RiCqY7 Decoupled weight decay regularization . In International Conference on Learning Representations

    2019

  9. [12]

    Maas, Raymond E

    Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. http://www.aclweb.org/anthology/P11-1015 Learning word vectors for sentiment analysis . In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142--150, Portland, Oregon, USA. Associa...

    2011

  10. [13]

    P. Malo, A. Sinha, P. Korhonen, J. Wallenius, and P. Takala. 2014. Good debt or bad debt: Detecting semantic orientations in economic texts. Journal of the Association for Information Science and Technology, 65

    2014

  11. [15]

    Anis Radianis. 2026. https://arxiv.org/abs/2605.19008 Learn-by-wire training control governance: Bounded autonomous training under stress for stability and efficiency . Preprint, arXiv:2605.19008

    Pith/arXiv arXiv 2026

  12. [18]

    Soheil Zibakhsh Shabgahi, Nojan Sheybani, Aiden Tabrizi, and Farinaz Koushanfar. 2024. https://arxiv.org/abs/2311.17279 Livetune: Dynamic parameter tuning for feedback-driven optimization . Preprint, arXiv:2311.17279

    Pith/arXiv arXiv 2024

  13. [19]

    Manning, Andrew Ng, and Christopher Potts

    Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. https://aclanthology.org/D13-1170/ Recursive deep models for semantic compositionality over a sentiment treebank . In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631--1642, Seattle, Washi...

    2013

  14. [20]

    Evan Pete Walsh, Luca Soldaini, Dirk Groeneveld, Kyle Lo, Shane Arora, Akshita Bhagia, Yuling Gu, Shengyi Huang, Matt Jordan, Nathan Lambert, Dustin Schwenk, Oyvind Tafjord, Taira Anderson, David Atkinson, Faeze Brahman, Christopher Clark, Pradeep Dasigi, Nouha Dziri, Allyson Ettinger, and 23 others. 2025. https://arxiv.org/abs/2501.00656 2 OLM o 2 furiou...

    Pith/arXiv arXiv 2025

  15. [22]

    Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer. 2022. https://arxiv.org/abs/2205.01068 Opt: Open pre-trained transformer language...

    Pith/arXiv arXiv 2022

  16. [23]

    Wentao Zhang, Yang Young Lu, and Yuntian Deng. 2025. https://doi.org/10.18653/v1/2025.emnlp-demos.65 Interactive training: Feedback-driven neural network optimization . In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 851--861, Suzhou, China. Association for Computational Linguistics

    work page doi:10.18653/v1/2025.emnlp-demos.65 2025

  17. [24]

    Interactive Training: Feedback-Driven Neural Network Optimization

    Zhang, Wentao and Lu, Yang Young and Deng, Yuntian. Interactive Training: Feedback-Driven Neural Network Optimization. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. 2025. doi:10.18653/v1/2025.emnlp-demos.65. arXiv:2510.02297

    arXiv 2025

  18. [25]

    2022 , eprint=

    OPT: Open Pre-trained Transformer Language Models , author=. 2022 , eprint=

    2022

  19. [26]

    Smith and Hannaneh Hajishirzi , booktitle=

    Evan Pete Walsh and Luca Soldaini and Dirk Groeneveld and Kyle Lo and Shane Arora and Akshita Bhagia and Yuling Gu and Shengyi Huang and Matt Jordan and Nathan Lambert and Dustin Schwenk and Oyvind Tafjord and Taira Anderson and David Atkinson and Faeze Brahman and Christopher Clark and Pradeep Dasigi and Nouha Dziri and Allyson Ettinger and Michal Guerqu...

    2025

  20. [27]

    2020 , note =

    Experiment Tracking with Weights and Biases , author =. 2020 , note =

    2020

  21. [28]

    2020 , month = jun, license =

    Arakelyan, Gor and Soghomonyan, Gevorg and. 2020 , month = jun, license =. doi:10.5281/zenodo.6536394 , note =

    work page doi:10.5281/zenodo.6536394 2020

  22. [29]

    Journal of Machine Learning Research , year =

    James Bergstra and Yoshua Bengio , title =. Journal of Machine Learning Research , year =

  23. [30]

    2017 , eprint=

    Population Based Training of Neural Networks , author=. 2017 , eprint=

    2017

  24. [31]

    Conference on Parsimony and Learning , pages =

    AgentHPO: Large Language Model Agent for Hyper-Parameter Optimization , author =. Conference on Parsimony and Learning , pages =. 2025 , editor =. 2402.01881 , archivePrefix =

    Pith/arXiv arXiv 2025

  25. [32]

    2024 , eprint=

    LiveTune: Dynamic Parameter Tuning for Feedback-Driven Optimization , author=. 2024 , eprint=

    2024

  26. [33]

    Journal of Artificial Intelligence Research , volume =

    Automated Dynamic Algorithm Configuration , author =. Journal of Artificial Intelligence Research , volume =. 2022 , doi =. 2205.13881 , archivePrefix =

    Pith/arXiv arXiv 2022

  27. [34]

    Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages =

    Optuna: A Next-generation Hyperparameter Optimization Framework , author =. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages =. 2019 , doi =. 1907.10902 , archivePrefix =

    Pith/arXiv arXiv 2019

  28. [35]

    When Is an

    Rodrigues, Carson and Vas, Oysturn and DCosta, Isaiah Abner and Prabhakaran, Nithish Kumar , year =. When Is an. 2606.21641 , archivePrefix =

    Pith/arXiv arXiv

  29. [36]

    2606.29871 , archivePrefix =

    Rao, Anjali and Advani, Nikhil Kamalkumar , year =. 2606.29871 , archivePrefix =

    Pith/arXiv arXiv

  30. [37]

    2026 , eprint =

    Learn-by-Wire Training Control Governance: Bounded Autonomous Training Under Stress for Stability and Efficiency , author =. 2026 , eprint =

    2026

  31. [38]

    2505.09388 , archivePrefix =

    Yang, An and Li, Anfeng and Yang, Baosong and Zhang, Beichen and Hui, Binyuan and Zheng, Bo and Yu, Bowen and Gao, Chang and Huang, Chengen and Lv, Chenxu and Zheng, Chujie and Liu, Dayiheng and Zhou, Fan and Huang, Fei and Hu, Feng and Ge, Hao and Wei, Haoran and Lin, Huan and Tang, Jialong and Yang, Jian and Tu, Jianhong and Zhang, Jianwei and Yang, Jia...

    Pith/arXiv arXiv

  32. [39]

    Penedo, Guilherme and Kydl. The. Advances in Neural Information Processing Systems , volume =. 2024 , publisher =. doi:10.52202/079017-0970 , eprint =

    doi:10.52202/079017-0970 2024

  33. [40]

    2024 , url =

    Muon: An Optimizer for Hidden Layers in Neural Networks , author =. 2024 , url =

    2024

  34. [41]

    and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle =

    Hu, Edward J. and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle =. 2022 , eprint =

    2022

  35. [42]

    Shao, Zhihong and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Song, Junxiao and Bi, Xiao and Zhang, Haowei and Zhang, Mingchuan and Li, Y. K. and Wu, Y. and Guo, Daya , year =. 2402.03300 , archivePrefix =

    Pith/arXiv arXiv

  36. [43]

    Lu, Chris and Lu, Cong and Lange, Robert Tjarko and Foerster, Jakob and Clune, Jeff and Ha, David , year =. The. 2408.06292 , archivePrefix =

    Pith/arXiv arXiv

  37. [44]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations , pages =

    Transformers: State-of-the-Art Natural Language Processing , author =. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations , pages =. 2020 , publisher =. doi:10.18653/v1/2020.emnlp-demos.6 , eprint =

    doi:10.18653/v1/2020.emnlp-demos.6 2020

  38. [45]

    International Conference on Learning Representations , year=

    Decoupled Weight Decay Regularization , author=. International Conference on Learning Representations , year=

  39. [46]

    Barbieri, Francesco and Camacho-Collados, Jose and Espinosa-Anke, Luis and Neves, Leonardo , booktitle=

  40. [47]

    Proceedings of the 7th ACM Conference on Recommender Systems , pages =

    McAuley, Julian and Leskovec, Jure , title =. Proceedings of the 7th ACM Conference on Recommender Systems , pages =. 2013 , isbn =. doi:10.1145/2507157.2507163 , abstract =

    arXiv 2013

  41. [48]

    Journal of the Association for Information Science and Technology , year=

    Good debt or bad debt: Detecting semantic orientations in economic texts , author=. Journal of the Association for Information Science and Technology , year=

  42. [49]

    and Daly, Raymond E

    Maas, Andrew L. and Daly, Raymond E. and Pham, Peter T. and Huang, Dan and Ng, Andrew Y. and Potts, Christopher , title =. Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies , month =. 2011 , address =

    2011

  43. [50]

    BERT : Pre-training of Deep Bidirectional Transformers for Language Understanding

    Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina. BERT : Pre-training of Deep Bidirectional Transformers for Language Understanding. Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). 2019. doi:10.18653/v...

    doi:10.18653/v1/n19-1423 2019

  44. [51]

    and Ng, Andrew and Potts, Christopher

    Socher, Richard and Perelygin, Alex and Wu, Jean and Chuang, Jason and Manning, Christopher D. and Ng, Andrew and Potts, Christopher. Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank. Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing. 2013

    2013

  45. [52]

    doi:10.57967/hf/2497 , publisher =

    Lozhkov, Anton and Ben Allal, Loubna and von Werra, Leandro and Wolf, Thomas , title =. doi:10.57967/hf/2497 , publisher =

    doi:10.57967/hf/2497