pith. sign in

arxiv: 2606.17660 · v1 · pith:3YMCGVB5new · submitted 2026-06-16 · 💻 cs.LG · cs.AI

TuneAhead: Predicting Fine-tuning Performance Before Full Training Begins

Yuxiang Luo , Haonan Long , Chen Wang , Qiqi Duan , Xiaotian Lin , Yanwei Xu , Yuyu Luo , Weikai Yang
show 1 more author
Nan Tang
This is my paper

Pith reviewed 2026-06-27 01:37 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords fine-tuning performance predictionlarge language modelsmeta-feature vectorsprobe featuresperformance estimationSHAP attributions
0
0 comments X

The pith

TUNEAHEAD predicts LLM fine-tuning performance from dataset features and a short probe run.

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

The paper tries to establish that final fine-tuning outcomes can be estimated before any full training begins. It does so by turning each possible run into a meta-feature vector built from fixed dataset statistics plus measurements collected during one brief standardized probe. A predictor then turns that vector into a performance number, and the whole system is checked against more than 1300 actual fine-tuning experiments on Qwen2.5-7B-Instruct. A sympathetic reader would care because fine-tuning is costly and can sometimes make models worse; early accurate forecasts could let practitioners avoid many wasted runs. The same features also supply explanations for why any given prediction is high or low.

Core claim

TUNEAHEAD encodes each candidate run as a meta-feature vector that combines static dataset descriptors with dynamic probe features from a short standardized probe. A predictor maps these features to performance estimates. Across 1,300+ fine-tuning runs on Qwen2.5-7B-Instruct, TUNEAHEAD consistently outperforms strong baselines such as Early-Stop Extrapolation and ProxyLM. On a held-out test set of 370 runs, TUNEAHEAD achieves an RMSE of 1.47 percentage points and places 95.1% of predictions within +3/-3 percentage points of the true score.

What carries the argument

The meta-feature vector that combines static dataset descriptors with dynamic features from a short standardized probe and feeds them to a predictor for performance estimates.

Load-bearing premise

Static dataset descriptors together with measurements from one short probe already contain enough information to forecast final performance no matter how the data quality or hyperparameters vary.

What would settle it

Applying the same predictor to a fresh collection of fine-tuning runs on different models or datasets and obtaining RMSE well above 1.47 percentage points would show the method does not generalize.

Figures

Figures reproduced from arXiv: 2606.17660 by Chen Wang, Haonan Long, Nan Tang, Qiqi Duan, Weikai Yang, Xiaotian Lin, Yanwei Xu, Yuxiang Luo, Yuyu Luo.

Figure 1
Figure 1. Figure 1: Predicting fine-tuning performance: (A) Without TUNEAHEAD: failed runs are only identified after training, wasting computational resources and time. (B) With TUNEAHEAD: low-cost features predict performance in advance, enabling go/no-go decisions and diagnosis for the failure cases. (A) Without TuneAhead 1 2 3 4 5 6 7 8 9 10 Total compute (no prediction) — 30h (B) With TuneAhead 1 2 4 7 9 Overhead + comput… view at source ↗
Figure 2
Figure 2. Figure 2: Compute time for 10 runs without TUNEAHEAD (A) vs. with TUNEAHEAD (B). With TUNEAHEAD (see [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: TUNEAHEAD Overview. Stage 1 (Meta-dataset curation) builds meta-feature vectors Vi,j by combining static features with dynamic features. Stage 2 (Predictive & Diagnostic Modeling) maps Vi,j to performance predictions and uses SHAP for diagnostics. tuning configuration. We therefore run a standardized 100- step probe for each dataset–hyperparameter pair (Di , Hj ). The resulting dynamic features should be i… view at source ↗
Figure 4
Figure 4. Figure 4: Predicted vs True accuracy across methods. The diagonal line (y=x) indicates a perfect prediction [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: (a) SHAP summary plot ranking the global importance of meta-features for predicting fine-tuning success; (b) SHAP waterfall plot for a representative failure case (the model correctly predicted low performance). contribution of TUNEAHEAD is its ability to provide diag￾nostic insights (G3). In this section, we use TreeSHAP to analyze the trained model and understand the key drivers of fine-tuning success or… view at source ↗
Figure 6
Figure 6. Figure 6: Effect of probe length on prediction accuracy, stability, and time cost. (a) Accuracy at 2pp steadily improves with longer probe runs but exhibits diminishing returns beyond 100 steps. (b) RMSE decreases sharply in the early stage and stabilizes after 100 steps. (c) Average probe time cost grows near-linearly with probe length, with 200 steps requiring about 1.5x the cost of 100 steps. more than 1,300 fine… view at source ↗
read the original abstract

Fine-tuning large language models (LLMs) is compute-intensive and error-prone: model performance depends sensitively on data quality and hyperparameter choices, and na\"ive runs can even degrade model performance. This raises a practical question:can we predict fine-tuning performance before committing to a full training run? We present TUNEAHEAD, a lightweight framework for pre-hoc prediction of fine-tuning performance. TUNEAHEAD encodes each candidate run as a meta-feature vector that combines static dataset descriptors with dynamic probe features from a short standardized probe. A predictor maps these features to performance estimates, while SHAP-based attributions provide interpretable diagnostics that reveal which specific features drive the prediction. Across 1,300+ fine-tuning runs on Qwen2.5-7B-Instruct, TUNEAHEAD consistently outperforms strong baselines such as Early-Stop Extrapolation and ProxyLM. On a held-out test set of 370 runs, TUNEAHEAD achieves an RMSE of 1.47 percentage points and places 95.1% of predictions within +3/-3 percentage points of the true score. These accurate continuous predictions support practical go/no-go screening policies that can reduce unnecessary full fine-tuning while retaining most promising runs.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces TUNEAHEAD, a lightweight framework for pre-hoc prediction of LLM fine-tuning performance. Each candidate run is encoded as a meta-feature vector combining static dataset descriptors with dynamic features extracted from a short standardized probe; a meta-predictor then maps these features to an estimated final performance score, accompanied by SHAP-based attributions. On more than 1,300 fine-tuning runs of Qwen2.5-7B-Instruct the method outperforms baselines such as Early-Stop Extrapolation and ProxyLM; on a held-out test set of 370 runs it reports RMSE = 1.47 percentage points and places 95.1 % of predictions inside a ±3 percentage-point band.

Significance. If the reported accuracy generalizes, the approach could materially reduce wasted compute by enabling early go/no-go screening of fine-tuning configurations. The evaluation scale (>1,300 runs) and the explicitly non-circular feature construction (dataset descriptors plus an independent short probe) are concrete strengths. The provision of SHAP diagnostics adds practical interpretability that is often absent from pure black-box predictors.

major comments (2)
  1. [Abstract] Abstract: the claim that predictions succeed 'across varied … hyperparameter choices without requiring the full training trajectory' is load-bearing, yet the probe is described as 'standardized' (fixed settings). If the dynamic features (early loss, gradient norms, etc.) are collected under a single fixed learning rate/batch size/optimizer, they cannot reflect hyperparameter-specific convergence behavior; the low RMSE on the held-out set would then be consistent with learning correlations inside the probe regime rather than true transfer across HP regimes.
  2. [Evaluation] Evaluation section (implied by the 370-run held-out test set): no information is given on how the train/test split was constructed with respect to hyperparameter diversity or dataset similarity. Without explicit stratification or leakage controls, the 1.47 RMSE and 95.1 % within-band statistic cannot be taken as evidence of generalization across the very hyperparameter variations the central claim asserts.
minor comments (1)
  1. [Abstract] Abstract contains a typographic artifact ('na"ive'); standard spelling is 'naive'.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback. The comments highlight important aspects of our claims and evaluation design. We address each major comment below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that predictions succeed 'across varied … hyperparameter choices without requiring the full training trajectory' is load-bearing, yet the probe is described as 'standardized' (fixed settings). If the dynamic features (early loss, gradient norms, etc.) are collected under a single fixed learning rate/batch size/optimizer, they cannot reflect hyperparameter-specific convergence behavior; the low RMSE on the held-out set would then be consistent with learning correlations inside the probe regime rather than true transfer across HP regimes.

    Authors: We appreciate the referee pointing out this distinction. The probe is run under fixed, standardized settings to ensure consistent and low-cost feature extraction that does not depend on the target run's hyperparameters. The meta-predictor is trained across a collection of runs that themselves use varied hyperparameters, so it learns to associate the resulting (dataset + fixed-probe) feature vectors with the final performance achieved under those specific hyperparameters. We agree, however, that the features themselves do not encode hyperparameter-specific dynamics. To prevent any overstatement of the generalization claim, we will revise the abstract to clarify that predictions are made for the hyperparameter configurations observed in the training data, using static descriptors plus a fixed probe, rather than implying direct transfer to arbitrary unseen hyperparameter regimes. revision: yes

  2. Referee: [Evaluation] Evaluation section (implied by the 370-run held-out test set): no information is given on how the train/test split was constructed with respect to hyperparameter diversity or dataset similarity. Without explicit stratification or leakage controls, the 1.47 RMSE and 95.1 % within-band statistic cannot be taken as evidence of generalization across the very hyperparameter variations the central claim asserts.

    Authors: The referee is correct that the manuscript currently lacks explicit details on the train/test split procedure with respect to hyperparameter diversity and dataset similarity. We will add a dedicated paragraph in the evaluation section describing how the 370-run held-out set was constructed, including any stratification by key hyperparameters (learning rate, batch size, etc.) and controls for dataset overlap or similarity to ensure the reported metrics reflect generalization across the hyperparameter variations present in the data. revision: yes

Circularity Check

0 steps flagged

No circularity: predictions use independent short-probe features

full rationale

The method constructs meta-feature vectors from static dataset descriptors plus dynamic signals extracted from a short standardized probe run; a separate predictor then maps these to performance estimates. No equation or step in the described pipeline reduces the output to a quantity defined by the full training trajectory itself. Evaluation on a held-out set of 370 runs is performed against actual full-run scores, preserving independence. No self-citation load-bearing steps or fitted-input-as-prediction patterns are present in the provided description.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The framework rests on the domain assumption that short probes capture predictive signal about full runs and that the chosen meta-features generalize; no free parameters or invented entities are explicitly introduced in the abstract.

free parameters (1)
  • meta-predictor parameters
    The performance predictor is trained on the collected fine-tuning runs, so its internal weights are fitted to data.
axioms (1)
  • domain assumption Short standardized probe features correlate sufficiently with full fine-tuning outcomes
    Invoked when the method uses probe data to stand in for complete training trajectories.

pith-pipeline@v0.9.1-grok · 5766 in / 1359 out tokens · 31814 ms · 2026-06-27T01:37:31.426136+00:00 · methodology

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

123 extracted references · 12 canonical work pages · 1 internal anchor

  1. [1]

    Scaling Learning Algorithms Towards

    Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards

  2. [2]

    and Osindero, Simon and Teh, Yee Whye , journal =

    Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume =

  3. [3]

    2016 , publisher=

    Deep learning , author=. 2016 , publisher=

    2016

  4. [4]

    2025 , eprint=

    COSMOS: Predictable and Cost-Effective Adaptation of LLMs , author=. 2025 , eprint=

    2025

  5. [5]

    2024 , eprint=

    Fine-Tuning or Fine-Failing? Debunking Performance Myths in Large Language Models , author=. 2024 , eprint=

    2024

  6. [6]

    2025 , eprint=

    LENSLLM: Unveiling Fine-Tuning Dynamics for LLM Selection , author=. 2025 , eprint=

    2025

  7. [7]

    2017 , eprint=

    A Unified Approach to Interpreting Model Predictions , author=. 2017 , eprint=

    2017

  8. [8]

    2020 , eprint=

    Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few Examples , author=. 2020 , eprint=

    2020

  9. [9]

    2021 , eprint=

    Dataset2Vec: Learning Dataset Meta-Features , author=. 2021 , eprint=

    2021

  10. [10]

    2022 , eprint=

    Model Zoos: A Dataset of Diverse Populations of Neural Network Models , author=. 2022 , eprint=

    2022

  11. [11]

    Proceedings of the 24th International Conference on Artificial Intelligence , pages =

    Domhan, Tobias and Springenberg, Jost Tobias and Hutter, Frank , title =. Proceedings of the 24th International Conference on Artificial Intelligence , pages =. 2015 , isbn =

    2015

  12. [12]

    2020 , eprint=

    Scaling Laws for Neural Language Models , author=. 2020 , eprint=

    2020

  13. [13]

    2024 , eprint=

    ProxyLM: Predicting Language Model Performance on Multilingual Tasks via Proxy Models , author=. 2024 , eprint=

    2024

  14. [14]

    2020 , eprint=

    Dataset Cartography: Mapping and Diagnosing Datasets with Training Dynamics , author=. 2020 , eprint=

    2020

  15. [15]

    2019 , eprint=

    Data Shapley: Equitable Valuation of Data for Machine Learning , author=. 2019 , eprint=

    2019

  16. [16]

    2025 , eprint=

    Refine-n-Judge: Curating High-Quality Preference Chains for LLM-Fine-Tuning , author=. 2025 , eprint=

    2025

  17. [17]

    2018 , eprint=

    The Dataset Nutrition Label: A Framework To Drive Higher Data Quality Standards , author=. 2018 , eprint=

    2018

  18. [18]

    2021 , eprint=

    MAUVE: Measuring the Gap Between Neural Text and Human Text using Divergence Frontiers , author=. 2021 , eprint=

    2021

  19. [19]

    2023 , eprint=

    GENTLE: A Genre-Diverse Multilayer Challenge Set for English NLP and Linguistic Evaluation , author=. 2023 , eprint=

    2023

  20. [20]

    2021 , eprint=

    On the Importance of Gradients for Detecting Distributional Shifts in the Wild , author=. 2021 , eprint=

    2021

  21. [21]

    2018 , eprint=

    GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks , author=. 2018 , eprint=

    2018

  22. [22]

    Louis , year=

    Shuofeng Zhang and Isaac Reid and Guillermo Valle-Perez and Ard A. Louis , year=

  23. [23]

    2024 , url=

    Gradient norm as a powerful proxy to out-of-distribution error estimation , author=. 2024 , url=

    2024

  24. [24]

    2023 , eprint=

    Gradient Norm Aware Minimization Seeks First-Order Flatness and Improves Generalization , author=. 2023 , eprint=

    2023

  25. [25]

    2021 , eprint=

    Self-Validation: Early Stopping for Single-Instance Deep Generative Priors , author=. 2021 , eprint=

    2021

  26. [26]

    2025 , eprint=

    Autoencoder-Based Framework to Capture Vocabulary Quality in NLP , author=. 2025 , eprint=

    2025

  27. [27]

    2025 , eprint=

    Massive Supervised Fine-tuning Experiments Reveal How Data, Layer, and Training Factors Shape LLM Alignment Quality , author=. 2025 , eprint=

    2025

  28. [28]

    2025 , eprint=

    Assessing the Impact of the Quality of Textual Data on Feature Representation and Machine Learning Models , author=. 2025 , eprint=

    2025

  29. [29]

    Model Explainability using SHAP Values for LightGBM Predictions , year=

    Bugaj, Michal and Wrobel, Krzysztof and Iwaniec, Joanna , booktitle=. Model Explainability using SHAP Values for LightGBM Predictions , year=

  30. [30]

    Proceedings of the 31st International Conference on Neural Information Processing Systems , pages =

    Ke, Guolin and Meng, Qi and Finley, Thomas and Wang, Taifeng and Chen, Wei and Ma, Weidong and Ye, Qiwei and Liu, Tie-Yan , title =. Proceedings of the 31st International Conference on Neural Information Processing Systems , pages =. 2017 , isbn =

    2017

  31. [31]

    2025 , issue_date =

    Wang, Ying and Wang, Pengxin and Tansey, Kevin and Liu, Junming and Delaney, Bethany and Quan, Wenting , title =. 2025 , issue_date =. doi:10.1016/j.compag.2024.109758 , journal =

    work page doi:10.1016/j.compag.2024.109758 2025

  32. [32]

    Garcia and Carlos Soares and Joaquin Vanschoren and André C.P.L.F

    Adriano Rivolli and Luís P.F. Garcia and Carlos Soares and Joaquin Vanschoren and André C.P.L.F. Meta-features for meta-learning , journal =. 2022 , issn =. doi:https://doi.org/10.1016/j.knosys.2021.108101 , url =

    work page doi:10.1016/j.knosys.2021.108101 2022

  33. [33]

    2022 , eprint=

    Predicting Fine-Tuning Performance with Probing , author=. 2022 , eprint=

    2022

  34. [34]

    2024 , eprint=

    Predictable Emergent Abilities of LLMs: Proxy Tasks Are All You Need , author=. 2024 , eprint=

    2024

  35. [35]

    2020 , eprint=

    The Break-Even Point on Optimization Trajectories of Deep Neural Networks , author=. 2020 , eprint=

    2020

  36. [36]

    2019 , eprint=

    Visualizing and Understanding the Effectiveness of BERT , author=. 2019 , eprint=

    2019

  37. [37]

    2023 , eprint=

    Efficient Bayesian Learning Curve Extrapolation using Prior-Data Fitted Networks , author=. 2023 , eprint=

    2023

  38. [38]

    International Conference on Learning Representations (ICLR) , year =

    Measuring Massive Multitask Language Understanding , author =. International Conference on Learning Representations (ICLR) , year =

  39. [39]

    Qwen2.5 Technical Report

    Qwen2.5 Technical Report , author =. arXiv preprint arXiv:2412.15115 , year =. doi:10.48550/arXiv.2412.15115 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2412.15115

  40. [40]

    2025 , url =

    Qwen2.5-7B-Instruct , author =. 2025 , url =

    2025

  41. [41]

    2025 , eprint=

    A Multi-Power Law for Loss Curve Prediction Across Learning Rate Schedules , author=. 2025 , eprint=

    2025

  42. [42]

    2020 , eprint=

    Don't Stop Pretraining: Adapt Language Models to Domains and Tasks , author=. 2020 , eprint=

    2020

  43. [43]

    Enhancing Data Quality through Simple De-duplication: Navigating Responsible Computational Social Science Research

    Mu, Yida and Jin, Mali and Song, Xingyi and Aletras, Nikolaos. Enhancing Data Quality through Simple De-duplication: Navigating Responsible Computational Social Science Research. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024. doi:10.18653/v1/2024.emnlp-main.694

    work page doi:10.18653/v1/2024.emnlp-main.694 2024

  44. [44]

    2025 , eprint=

    Measuring Data Diversity for Instruction Tuning: A Systematic Analysis and A Reliable Metric , author=. 2025 , eprint=

    2025

  45. [45]

    2023 , eprint=

    The Vendi Score: A Diversity Evaluation Metric for Machine Learning , author=. 2023 , eprint=

    2023

  46. [46]

    2018 , eprint=

    Visualizing the Loss Landscape of Neural Nets , author=. 2018 , eprint=

    2018

  47. [47]

    2020 , eprint=

    Adversarial Weight Perturbation Helps Robust Generalization , author=. 2020 , eprint=

    2020

  48. [48]

    Mathematics , VOLUME =

    Liang, Hailun and Zheng, Haowen and Wang, Hao and He, Liu and Lin, Haoyi and Liang, Yanyan , TITLE =. Mathematics , VOLUME =. 2025 , NUMBER =

    2025

  49. [49]

    Technometrics , volume =

    Ridge Regression: Biased Estimation for Nonorthogonal Problems , author =. Technometrics , volume =. 1970 , publisher =

    1970

  50. [50]

    Advances in Neural Information Processing Systems , volume =

    Support Vector Regression Machines , author =. Advances in Neural Information Processing Systems , volume =. 1997 , publisher =

    1997

  51. [51]

    Statistics and Computing , volume =

    A Tutorial on Support Vector Regression , author =. Statistics and Computing , volume =. 2004 , doi =

    2004

  52. [52]

    Machine Learning , volume =

    Random Forests , author =. Machine Learning , volume =. 2001 , doi =

    2001

  53. [53]

    Nature , volume =

    Learning Representations by Back-Propagating Errors , author =. Nature , volume =. 1986 , doi =

    1986

  54. [54]

    Neural Networks , volume =

    Multilayer Feedforward Networks Are Universal Approximators , author =. Neural Networks , volume =. 1989 , doi =

    1989

  55. [55]

    Journal of Machine Learning Research , volume =

    Scikit-learn: Machine Learning in Python , author =. Journal of Machine Learning Research , volume =. 2011 , url =

    2011

  56. [56]

    arXiv preprint arXiv:1706.10239 , year=

    Towards Understanding Generalization of Deep Learning: Perspective of Loss Landscapes , author=. arXiv preprint arXiv:1706.10239 , year=

    Pith/arXiv arXiv

  57. [57]

    arXiv preprint arXiv:2412.13573 , year=

    Seeking Consistent Flat Minima for Better Domain Generalization via Refining Loss Landscapes , author=. arXiv preprint arXiv:2412.13573 , year=

    arXiv

  58. [58]

    2019 , eprint=

    Characterizing classification datasets: a study of meta-features for meta-learning , author=. 2019 , eprint=

    2019

  59. [59]

    2024 , eprint=

    Machine Translation Meta Evaluation through Translation Accuracy Challenge Sets , author=. 2024 , eprint=

    2024

  60. [60]

    2023 , eprint=

    Advances and Challenges in Meta-Learning: A Technical Review , author=. 2023 , eprint=

    2023

  61. [61]

    2018 , eprint=

    Don't Give Me the Details, Just the Summary! Topic-Aware Convolutional Neural Networks for Extreme Summarization , author=. 2018 , eprint=

    2018

  62. [62]

    ROUGE : A Package for Automatic Evaluation of Summaries

    Lin, Chin-Yew. ROUGE : A Package for Automatic Evaluation of Summaries. Text Summarization Branches Out. 2004

    2004

  63. [63]

    2018 , eprint=

    A Survey of Machine Learning for Big Code and Naturalness , author=. 2018 , eprint=

    2018

  64. [64]

    , title =

    Raffel, Colin and Shazeer, Noam and Roberts, Adam and Lee, Katherine and Narang, Sharan and Matena, Michael and Zhou, Yanqi and Li, Wei and Liu, Peter J. , title =. J. Mach. Learn. Res. , month = jan, articleno =. 2020 , issue_date =

    2020

  65. [65]

    2022 , eprint=

    Deduplicating Training Data Makes Language Models Better , author=. 2022 , eprint=

    2022

  66. [66]

    2023 , eprint=

    Extracting Training Data from Diffusion Models , author=. 2023 , eprint=

    2023

  67. [67]

    2022 , eprint=

    Training Compute-Optimal Large Language Models , author=. 2022 , eprint=

    2022

  68. [68]

    2019 , eprint=

    Deep Anomaly Detection with Outlier Exposure , author=. 2019 , eprint=

    2019

  69. [69]

    2022 , eprint=

    Confident Learning: Estimating Uncertainty in Dataset Labels , author=. 2022 , eprint=

    2022

  70. [70]

    2021 , eprint=

    Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpus , author=. 2021 , eprint=

    2021

  71. [71]

    2020 , eprint=

    The Curious Case of Neural Text Degeneration , author=. 2020 , eprint=

    2020

  72. [72]

    A Qualitative Comparison of C o QA , SQ u AD 2.0 and Q u AC

    Yatskar, Mark. A Qualitative Comparison of C o QA , SQ u AD 2.0 and Q u AC. 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/v1/N19-1241

    work page doi:10.18653/v1/n19-1241 2019

  73. [73]

    2016 , eprint=

    Exploring the Limits of Language Modeling , author=. 2016 , eprint=

    2016

  74. [74]

    2018 , eprint=

    Universal Sentence Encoder , author=. 2018 , eprint=

    2018

  75. [75]

    2020 , eprint=

    ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators , author=. 2020 , eprint=

    2020

  76. [76]

    2016 , eprint=

    A Diversity-Promoting Objective Function for Neural Conversation Models , author=. 2016 , eprint=

    2016

  77. [77]

    2020 , eprint=

    Fine-Tuning Language Models from Human Preferences , author=. 2020 , eprint=

    2020

  78. [78]

    2020 , eprint=

    Unsupervised Domain Clusters in Pretrained Language Models , author=. 2020 , eprint=

    2020

  79. [79]

    Domain Adaptation via Pseudo In-Domain Data Selection

    Axelrod, Amittai and He, Xiaodong and Gao, Jianfeng. Domain Adaptation via Pseudo In-Domain Data Selection. Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing. 2011

    2011

  80. [80]

    Khasentino, J., Belyaeva, A., Liu, X., Yang, Z., Furlotte, N

    Ji, Ziwei and Lee, Nayeon and Frieske, Rita and Yu, Tiezheng and Su, Dan and Xu, Yan and Ishii, Etsuko and Bang, Ye Jin and Madotto, Andrea and Fung, Pascale , year=. Survey of Hallucination in Natural Language Generation , volume=. ACM Computing Surveys , publisher=. doi:10.1145/3571730 , number=

    work page doi:10.1145/3571730

Showing first 80 references.