PaperlayerView exact revision on arXiv

Source-linked AI summary

MoMaTune: Model, Match, Then Tune Language Models for Numerical Records

Weimin Lyu, Jingfen Zhang, Lingdao Sha, Peng sun, Xiaowei Chen, Vincent Gao, Rui Song

arXiv:2610.15235v1cs.AI

TL;DR

Numerical records leave digit-position magnitude and cross-field relationships weakly supervised, raising whether causal decoders can learn them from serialized records. MoMaTune first adapts models on label-free records, optionally aligning partial views, then fine-tunes for target prediction. The two-stage recipes outperform Tune alone across tested model sizes, with reported gains in numerical completion and digit sensitivity.

  • Problem

    Serialized numerical records leave magnitude within values and dependencies across fields implicit, while prior work leaves open whether standard causal decoders can learn both directly from label-free records.

    Source: §1 · p. 1

  • Method

    MoMaTune applies dense next-token prediction to label-free records, optionally aligns same-record partial-view representations, then uses supervised fine-tuning for target prediction.

    Source: §1 · p. 2, §3.2 · p. 4

  • Results

    Across 1.5B, 4B, 9.6B, and 27B models, field-permuted MODEL+MATCH→TUNE outperforms TUNE alone by 1.0–2.6% on target-prediction scores.

    Source: §4.3 · p. 9

  • Takeaways & Limitations

    The authors recommend MODEL→TUNE as the core path and adding MATCH only when partial-view robustness is needed, since MATCH can reduce prediction quality for complementary views.

    Source: §4.2.2 · p. 8, §5 · p. 9

    The main results are almost entirely numerical, and TUNE’s target supervision limits conclusions to post-training numerical records and target specialization rather than broad tabular or label-free generality.

    Source: A LIMITATIONS · p. 12

Abstract

from arXiv · show

Despite remarkable progress in language understanding and generation, decoder language models continue to exhibit striking failures on numerical data. Numerical examples are typically serialized as text by concatenating feature names with their values (e.g., fat:3.4), while conventional task adaptation applies loss only to the downstream prediction target. This leaves two structures weakly supervised: within-value sensitivity, where digit position determines numerical magnitude, because token cross-entropy contains no explicit numerical metric; and across-field structure, where measurements in the same record are statistically dependent, because target-only tuning provides no direct loss on the measurements or their dependencies. We introduce MoMaTune (Model, Match, then Tune), a two-stage post-training strategy organized around these supervision gaps. Stage 1 applies Model, ordinary next-token prediction densely over label-free serialized records, and optionally adds Match, a same-record consistency objective that aligns representations of different partial views of the same example. Stage 2 applies Tune, supervised fine-tuning for downstream target prediction. We evaluate MoMaTune on 169,285 records from eight public datasets, with 339 of 340 input fields numerical. At 1.5B parameters, Model raises the eight-dataset masked-value imputation aggregate from 0.106 to 0.401, while the three-dataset correct-context gain rises from 0.058 to 0.167. On diagnostics, Model also increases sensitivity to numerically consequential digit changes, while Match improves robustness to informative partial views. Adding Tune yields higher target-supervised embedding aggregates for each tested Stage-1 parent, and the two-stage recipes obtain higher aggregate scores than Tune alone. The same pattern is observed at 4B, 9.6B, and 27B parameters when model embeddings are used for downstream prediction.

1 INTRODUCTION

Numerical records encode magnitude and relationships among measurements only implicitly, leaving open whether standard causal decoders can learn both from serialized records. MoMaTune adapts models to label-free records before target tuning, with optional consistency training for partial-view robustness.

1 INTRODUCTION

  • Numerical records leave digit-position magnitude and cross-field dependencies implicit, while prior approaches leave unresolved whether standard causal decoders can learn both structures.The paper frames these as within-value sensitivity and across-field structure.Source: §1 · p. 1
  • A first-versus-fourth fractional-digit edit changes 0.4271 by effects differing 1000×, while the frozen 1.5B model’s sensitivity ratio is only 2.08×.The ratio compares pooled-record embedding movement after first- and fourth-digit edits.Source: §1 · p. 2, Figure 1 · p. 2
  • Across eight datasets, the frozen 1.5B model’s masked-value imputation aggregate is 0.106, below the 0.212 field-conditional median baseline; its three-dataset context gain is 0.058.These diagnostics respectively probe completion and use of cross-field context.Source: §1 · p. 2
  • MoMaTune’s core path applies dense next-token prediction to label-free numerical records, then target-supervised fine-tuning; optional MATCH aligns partial views for robustness.MATCH’s effect depends on the partial-view construction.Source: §1 · p. 2
  • Evaluation spans embedding and generation tasks on eight public datasets with 169,285 records, including 339 numerical input fields among 340.Seven datasets contain only numerical input fields.Source: §1 · p. 3

2 RELATED WORK

Prior work applies language models to serialized records and develops methods to encode numerical or feature structure more explicitly. This paper asks whether a standard causal decoder can learn numerical magnitude and cross-field structure directly from label-free serialized records.

2 RELATED WORK

  • Numerical-structure methods modify tokenization, use feature-aware pretraining or subset views, or apply target-distance objectives, whereas this work retains causal decoding and unchanged tokenization.The cited approaches include Golkar et al. (2023), Yan et al. (2024), Ucar et al. (2021), and Zha et al. (2023).Source: §2 · p. 3
  • Decoder-derived encoders often use supervised pairs, modified attention, specialized pooling, or embedding-focused training; this work instead improves numerical representations while preserving causal decoding.The related approaches include E5-Mistral, LLM2Vec, and NV-Embed [Wang et al., 2024; BehnamGhader et al., 2024; Lee et al., 2025].Source: §2 · p. 3
  • Serialized-record methods establish language models’ usefulness for prediction or generation, but do not directly study what numerical structure causal decoders acquire through label-free record modeling before target supervision.Examples include TabLLM, GReaT, GTL, and TabuLa [Hegselmann et al., 2023; Borisov et al., 2022; Wen et al., 2024; Wang & Fritz, 2024].Source: §2 · p. 3

3 MOMATUNE: A POST-TRAINING FRAMEWORK FOR NUMERICAL RECORDS

MoMaTune represents numerical records as serialized key–value fields and uses label-free record adaptation before target-specializing fine-tuning. Its optional same-record consistency objective supports partial-view robustness, while field-order permutation diversifies record presentations.

3.1 DEFINITION: RECORDS, SERIALIZATION, AND EMBEDDINGS

  • A record contains measured key–value fields, and serialization maps their ordered sequence to text; the downstream target is excluded from Stage 1.The paper denotes the serialized record x(r), its embedding z(x), and the target y.Source: §3.1 · pp. 3–4

3.2 STAGE 1: RECORD ADAPTATION

  • MODEL applies ordinary causal next-token prediction to every visible token in a truncated, label-free record, supplying dense in-domain supervision without changing the loss geometry.MODEL alone minimizes record loss; MATCH is optional rather than required for every Stage-1 run.Source: §3.2 · p. 4, Figure 2 · p. 4
  • MATCH aligns embeddings of two label-free views of the same record; views may be disjoint field subsets or independently permuted and truncated prefixes.The permuted-prefix views can differ in both field order and visible content.Source: §3.2 · p. 4
  • Field-order permutation reorders complete key–value fields while preserving names and values, presenting measurements in varied contexts rather than fixed positions.This data augmentation is intended to encourage learning field identities and cross-field relationships without relying on position.Source: §3.2 · pp. 4–5

3.3 STAGE 2: TARGET SPECIALIZATION

  • TUNE specializes the Stage-1 checkpoint using a fixed-order record and target question, masking record and question positions so only answer tokens contribute to loss.Because TUNE observes the target, its embedding score measures supervised target specialization rather than label-free competence.Source: §3.3 · p. 5, Figure 3 · p. 5

4 EXPERIMENTAL DESIGN

The experiments assess whether post-training improves target prediction and numerical-record capabilities across mostly numerical datasets and model scales. MODEL improves target prediction and numerical completion; MATCH helps with compatible partial views, while TUNE specializes representations for target prediction.

4.1 EXPERIMENTAL SETTINGS

  • The evaluation covers 169,285 records across eight public datasets, with 339 of 340 input fields numerical, using embeddings for target prediction and masked-value generation for completion.The size sweep uses different checkpoint families, so it tests robustness across both scale and model family.Source: §4.1 · pp. 5–6

4.2.1 OVERALL PERFORMANCE: TARGET PREDICTION AND NUMERICAL COMPLETION

  • 0.753 EMB-8 is the highest reported aggregate, shared by field-permuted MODEL→TUNE and field-permuted MODEL+MATCH→TUNE, versus 0.688 for the frozen checkpoint.At 1.5B, fixed-order MODEL→TUNE reaches 0.736 EMB-8.Source: Table 1 · p. 6, §4.2.1 · p. 6
  • 0.106 to 0.448: Stage-1 raises mean masked-value completion across eight datasets; adding MATCH does not improve completion.The task asks the model to generate a held-out value and counts predictions within 10% as correct.Source: §4.2.1 · pp. 6–7, Table 2 · p. 7

4.2.2 MODEL, MATCH, AND TUNE CONTRIBUTE DISTINCT CAPABILITIES

  • MODEL raises correct-context gain from 0.058 to 0.167 and digit-sensitivity ratio from 2.08× to 13.83×, with field-permuted MODEL reaching 0.271 and 17.82×, respectively.The context gain compares completion with correct versus mismatched records; the sensitivity ratio compares embedding changes after first- versus fourth-digit edits.Source: §4.2.2 · p. 7, Table 3 · p. 8
  • For permuted-prefix views, MATCH reduces embedding displacement 15.3× and raises incomplete-view ROC–AUC from 0.719 to 0.758, but can lower AUC when paired views hold complementary evidence.With disjoint-half views, AUC falls from 0.584 to 0.533 after MATCH.Source: §4.2.2 · p. 8, Figure 4 · p. 8
  • TUNE raises EMB-8 by 0.8–1.9% over each tested 1.5B Stage-1 parent; across 1.5B–27B, MoMaTune beats TUNE alone by 1.0–2.6%.Masked-value imputation decreases for every tested parent after TUNE.Source: §4.2.2 · p. 8, Table 4 · p. 9
  • At 27B, field-permuted MODEL raises three-dataset mean imputation by 0.134 over the equal-update fixed-order control and improves every dataset.At 1.5B, field permutation also raises EMB-8 from 0.743 to 0.753 and context gain from 0.167 to 0.271.Source: §4.2.2 · pp. 8–9, Figure 5 · p. 9

4.3 ABLATION STUDIES

  • Across 1.5B, 4B, 9.6B, and 27B, field-permuted MODEL+MATCH→TUNE outperforms TUNE alone by 1.0–2.6%.Trends are consistent on two non-numerical datasets, including one with up to 95.3% non-numeric data; the default MODEL–MATCH loss weighting has the highest R2, with tested λ2 values within 0.005.Source: §4.3 · p. 9

5 CONCLUSION

The conclusion assigns MODEL the foundation role, TUNE downstream specialization, and MATCH optional partial-view robustness. The other statements address AI use, ethics, and reproducibility.

5 CONCLUSION

  • MODEL adapts the decoder to numerical-record structure, TUNE specializes representations for target prediction while reducing completion ability, and MATCH is optional for partial-view robustness.MATCH can reduce prediction quality when paired views contain complementary information; the authors recommend validating the intended view construction.Source: §5 · p. 9

AI USE STATEMENT

  • Generative AI assisted with literature organization, code, experiment bookkeeping, and manuscript polishing; authors reviewed the work and checked citations and numerical results.The authors state that results came from executed code and were checked against stored evaluation artifacts and logs.Source: AI USE STATEMENT · p. 10

ETHICS STATEMENT

  • The study uses public benchmark data without human participants or sensitive personal data, and any consequential future use would require domain-specific validation and oversight.Source: ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

  • The supplementary materials document dataset construction, methods, full results, protocols, and scripts for regenerating reported tables.They include result files and scripts alongside detailed specifications for recipes, evaluation, and model sizes.Source: REPRODUCIBILITY STATEMENT · p. 10

A LIMITATIONS

The study’s scope is bounded by its model-size comparisons, largely numerical benchmark, and target-supervised TUNE stage.

A LIMITATIONS

  • Changing architecture and tokenizer vocabulary across model sizes means the sweep does not establish a fitted scaling law.The mixed-type evidence bounds rather than settles that question; the main results remain almost entirely numerical, and TUNE observes targets while reducing imputation.Source: A LIMITATIONS · p. 12

B INTUITION: LLM NUMERACY ISSUE

The section explains why digit position matters numerically and tests whether pretrained decoders’ representations reflect that difference. The surveyed models show only modest leading-digit sensitivity relative to the input-value contrast.

B INTUITION: LLM NUMERACY ISSUE

  • A one-digit edit changes one token under both described conventions, although its numerical effect depends sharply on digit position.For 0.4271, changing the first decimal digit changes the value by 23.4%, while changing the fourth changes it by 0.0234%.Source: B INTUITION: LLM NUMERACY ISSUE · p. 12
  • The diagnostic compares embedding movement after editing the first versus fourth significant digit; it does not expect the embedding-distance ratio to equal the 1000× input contrast.The reference edits change 0.4271 to 0.5271 and 0.4272, respectively.Source: B INTUITION: LLM NUMERACY ISSUE · p. 12
  • Across seven surveyed models, digit-edit ratios are 1.12–2.14 against a 1000× input contrast, with the two tokenization conventions overlapping completely.The diagnostic suggests digit-blindness across the surveyed pretrained decoders, rather than a pattern isolated to one tokenizer or model family.Source: B INTUITION: LLM NUMERACY ISSUE · p. 12, Table 5 · p. 13

C.1 DATASET CONSTRUCTION

The section describes a predominantly numerical benchmark, two evaluation axes, and the model and training arms used to assess MoMaTune. Its supplied results include an EMB-8 comparison of TUNE presentation choices.

C.1 DATASET CONSTRUCTION

  • The public benchmark spans eight datasets: seven are numerical-only, and 339 of 340 input fields are numerical; 95.3% of records come from numerical-only datasets.Each OpenML dataset is converted to a float feature matrix, capped at 30,000 records where necessary, randomly permuted with NumPy seed 42, and split 80/10/10. The downstream target is stored separately and is absent from every Stage-1 record-adaptation serialization.Source: §C.1 · p. 13, Table 6 · p. 13

C.2 EVALUATION DIMENSIONS AND METRICS

  • Evaluation separates frozen-embedding probes, which fit supervised linear readouts, from direct decoder tests of generation and record-content tasks.The generation tasks include masked-value imputation, field lookup, held-out likelihood, and target emission; target emission is limited to prompts fitting a 496-token budget.Source: §C.2 · pp. 14, 16
  • EMB-8 averages two classification ROC-AUC values and six regression R2 values, while GEN-8 averages within-10% imputation accuracy across eight datasets.These unweighted aggregates summarize different metrics and do not replace per-dataset results; target-generation metrics are reported separately.Source: §C.2 · pp. 16–17

D.1 MODEL ARCHITECTURES

  • The checkpoint set includes Qwen3.5 and Qwen3.8 hybrid decoders whose vision components are unused; experiments operate on their text decoders.The cited model-checkpoint table identifies vocabulary size as a checkpoint characteristic.Source: §D.1 · p. 17, Table 9 · p. 17

D.2 COMPLETE ARM REGISTRY

  • The principal arms combine MODEL next-token record adaptation, optional MATCH consistency, and TUNE supervised fine-tuning; MODEL recipes use loss weight 0.5 and truncate serializations at 512 tokens.The complete registry also includes TUNE descendants of consistency-only and order-aware arms.Source: Table 10 · p. 18, §D.2 · p. 18
  • At 1.5B, standard presentation improves EMB-8 by 0.012/0.014 over field-permuted TUNE for the two field-permuted Stage-1 parents.The comparison isolates the supervised presentation choice; the MODEL+MATCH→TUNE comparison against MODEL→TUNE exceeds the embedding duplicate band, while aligned larger-scale aggregate effects remain small.Source: §D.2 · p. 18

D.3 TRAINING DETAILS

The section specifies the standard training controls and reports extended results, including a loss-weight sweep showing that MATCH’s effect varies by dataset. It also documents the per-dataset metrics and the scope of the reported controls.

D.3 TRAINING DETAILS

  • The standard Stage-1 arms use two epochs, effective batch 16, learning rate 10−5, maximum length 512, and seed 42, with FSDP for larger models.Other shared settings include AdamW, 100 warmup steps, cosine decay, and bfloat16.Source: §D.3 · p. 18
  • The 512-token budget truncates every MiniBooNE test row, exposing 35.8 of 50 complete fields on average; 91.5% of raw truncations end inside the next key-value unit.The rows contain 590–726 tokens, with a mean of 708.5.Source: §D.3 · p. 18

E.1.1 SENSITIVITY TO THE LOSS WEIGHTS

  • On Protein, MODEL+MATCH reaches R2 .420 at λ2=1.0, versus .396 at λ2=0 and .414 at λ2=4.0; on USDA, settings are within measurement spread.The Protein setting is within .005 of its best value for λ2 ∈[0.5, 2.0], and held-out Protein perplexity remains 2.13–2.14 across rows.Source: §E.1.1 · p. 19
  • The loss-weight controls support using λ2/λ1 as the operative hyperparameter: doubling both coefficients reproduces the reported row within .001, while removing Lrecord collapses the representation.With Lrecord removed, Protein R2 is .248 and perplexity is 92.6; the sweep covers two regression datasets, one scale, and one seed.Source: §E.1.1 · pp. 19–20

E.2 COMPLETE PER-DATASET TARGET-PREDICTION RESULTS

  • The detailed results report classification ACC and ROC–AUC, regression R2, MAE, and Pearson r, with higher values favorable except for MAE.The classification mean is the arithmetic mean of four displayed metrics; regression means are arithmetic means over six datasets, with MAE’s units differing across datasets.Source: Table 14 · p. 20, Table 15 · p. 21, Table 16 · p. 22

E.3 GENERALITY TO RECORDS WITH SUBSTANTIAL TEXT AND CATEGORICAL FIELDS

The additional mixed-type datasets indicate that the recipes tolerate non-numeric fields, while the ability to distinguish recipe effects depends on task headroom. The section also reports a scoped diagnostic of digit-position sensitivity in MiniBooNE.

E.3 GENERALITY TO RECORDS WITH SUBSTANTIAL TEXT AND CATEGORICAL FIELDS

  • The results support tolerance to text and categorical fields, but the paper bounds this conclusion to two datasets, one backbone, and one seed.On Census-KDD, recipes stay at or above the frozen baseline, but differences among recipes are not resolvable; NYC-Taxi shows the ordering with more headroom.Source: §E.3 · pp. 23–24

F COMPONENT EVIDENCE AND MECHANISM DIAGNOSTICS

  • The diagnostic intervals use paired record-bootstrap resamples and quantify uncertainty conditional on trained checkpoints, not variation across training seeds.The protocol uses 10,000 paired resamples and recomputes ROC–AUC while holding the fitted probe fixed.Source: F COMPONENT EVIDENCE AND MECHANISM… · p. 24

F.1 MODEL: NUMERICAL SENSITIVITY AND CROSS-FIELD CONTEXT

  • MODEL produces the largest increase in first-digit sensitivity over the frozen checkpoint, while MATCH contracts the proxy for both fixed-order and field-permuted MODEL.Every edited unit is complete in the encoded input, and every variant changes one tokenizer token.Source: Figure 6 · p. 25

F.1.1 DIGIT-POSITION SENSITIVITY PROTOCOL AND FULL RESULTS

  • The digit-position diagnostic uses 200 MiniBooNE records and complete encoded key-value units; every retained perturbation changes exactly one tokenizer token.It compares embedding-distance responses to changes in the first four significant digits and treats the distance ratio as exploratory because its scale is denominator-sensitive.Source: §F.1.1 · p. 25, Table 19 · p. 26

F.2 MATCH: STABILITY AND INCOMPLETE-RECORD PREDICTION

MATCH contracts partial-view displacement, but its predictive value depends on how the views are constructed. TUNE improves target prediction while reducing masked-value completion, and field-order permutation changes which fields are visible under the token budget.

F.2 MATCH: STABILITY AND INCOMPLETE-RECORD PREDICTION

  • With permuted-prefix training, MATCH reduces full-to-incomplete displacement 15.305× and raises mean incomplete-view AUC by 0.0392; with disjoint-half views, incomplete-view AUC falls by 0.0510.For permuted-prefix views, full-view AUC difference is unresolved at +0.00145; the authors suggest overlapping conditional information may explain why the view construction matters.Source: §F.2 · p. 26, Figure 7 · p. 27
  • The experiment supports a view-compatibility hypothesis, not a general rule for constructing view pairs; MATCH is presented as an optional robustness regularizer rather than a complete-record accuracy stage.Matched field-permuted recipes reach the same EMB-8 after TUNE, while Stage-1 completion changes from 0.448 to 0.445.Source: §F.2 · p. 26

F.3 TUNE: TARGET SPECIALIZATION AND COMPLETION TRADE-OFF

  • Returning to fixed field order during TUNE raises EMB-8 from 0.741 to 0.753 for MODEL and from 0.740 to 0.753 for model-and-match.The same section reports that the field-permuted MODEL→TUNE and model-and-match recipes reach 0.753 when TUNE uses fixed order.Source: §F.3 · p. 28
  • TUNE improves the primary target-prediction metric for every tested 1.5B Stage-1 parent, but masked-value completion falls from 0.401, 0.401, 0.448, and 0.445 to 0.145, 0.145, 0.221, and 0.234.Applications needing both target prediction and general record completion should retain the Stage-1 checkpoint or use a retention-aware multi-task objective.Source: §F.3 · p. 28

F.4 FIELD-ORDER PERMUTATION AND COMPUTE CONTROLS

  • Field-order permutation changes visible field subsets and local contexts under the fixed token budget; it is a data intervention, not an order-invariance test.The paired-view audit reports that permuted-prefix views change both field order and visible content.Source: §F.4 · p. 28

G.1 MODEL-SIZE ROBUSTNESS FROM 1.5B TO 27B

Across the four tested checkpoints, MODEL improves embedding and masked-value-imputation aggregates over frozen checkpoints, and the complete recipe exceeds TUNE alone on target prediction. MATCH does not yield scale-growing aggregate accuracy, while TUNE’s specialization trades off with record completion.

G.1 MODEL-SIZE ROBUSTNESS FROM 1.5B TO 27B

  • From 1.5B to 27B, fixed-order MODEL improves both the mean downstream metric and mean within-10% masked-value imputation over frozen checkpoints.The model-size figure is a one-seed robustness sweep, not a fitted scaling law, because checkpoint architectures and tokenizers differ.Source: Figure 8 · p. 29, §G.1 · p. 29

G.2 MASKED-VALUE COMPLETION ACROSS MODEL SIZES

  • Only frozen decoders and MODEL improve monotonically with scale in masked-value completion; every TUNE recipe remains below MODEL alone.The paper attributes this pattern to specializing the decoder to emit a single target rather than model the whole record.Source: §G.2 · pp. 30–31, Table 25 · p. 31

G.3 COMPLETE 27B RECIPE GRID

  • Across all four tested checkpoints, the complete recipe scores higher on target prediction than TUNE alone and reaches the highest reported 27B embedding score, 0.942.At equal update budgets, field-permuted MODEL exceeds extended-compute MODEL at 27B by 0.016 on embedding and 0.134 on imputation; adding MATCH is aggregate-inert or negative through 27B.Source: §G.3 · pp. 31–32, Table 26 · p. 31

H ADDITIONAL DIRECT-DECODER EVALUATIONS

Prompt-matched target emission is a secondary diagnostic on four datasets that fit the TUNE prompt budget. Its recipe rankings depend on the evaluation prompt, and four longer datasets are excluded to avoid a train/test prompt mismatch.

H.1 EVALUATION SCOPE AND PROMPT COMPATIBILITY

  • Target emission is a prompt-conditioned secondary diagnostic, distinct from the eight-dataset masked-value-imputation evaluation.The reported decoder tests use a 320-token cap for held-out record perplexity, masked-value imputation, and field lookup; target emission uses the TUNE template and a 496-token cap.Source: §H.1 · p. 32, §H.2 · p. 32

H.2 PROMPT-MATCHED TARGET EMISSION

  • Under the TUNE-matched prompt, all four TUNE recipes beat their task-specific floors on all four datasets, but no recipe dominates.MODEL+MATCH→TUNE is best on HIGGS, CPU, and Protein; the field-permuted complete recipe is best on Naval.Source: Table 27 · p. 32, §H.2 · p. 32
  • Four datasets exceed the 496-token TUNE budget, so right truncation removes the question-and-answer suffix and those datasets are excluded.The decoder aggregate in the main paper is eight-dataset masked-value imputation; target emission is reported separately for the four prompt-valid datasets.Source: §H.2 · p. 32, §H.1 · p. 32
  • Under the original 320-token instruction-style prompt, rankings change: TUNE only wins HIGGS, while the field-permuted complete recipe wins CPU, Naval, and Protein.All recipes beat floors on HIGGS, Naval, and Protein, but TUNE only and MODEL→TUNE miss the CPU floor.Source: §H.2 · p. 32

Related papers

Loading 2610.15235v1…