This dataset contains all mechanistic profiling data, Q-scores, and validation results from the paper "Training Quality Determines Efficiency Boundaries in Test-Time Reasoning". It covers 19 reasoning-specialized language models (1.5B–685B parameters) evaluated across four reasoning benchmarks (GSM8K, BoolQ, ARC-Easy, GPQA).

The central contribution is the Q-score framework — a composite training quality metric computed entirely from inference-time output properties that explains 61.9% of variance in token efficiency (r = +0.787, P = 0.001) across all models, independent of parameter count.

Dataset Structure

dataset/
├── mechanistic/
│   ├── mechanistic_dimensions_all_models.csv    # Core mechanistic data — 19 models × 4 datasets
│   └── ablation_training_variants.csv           # Qwen2.5-7B training ablation study
├── qscores/
│   ├── token_efficiency_all_models.csv          # Q-scores + efficiency metrics — all 19 models
│   └── qscore_cross_dataset.csv                 # Q-scores per model per dataset (pivot-ready)
├── validation/
│   ├── validation_framework_original13.csv      # Train/test split — original 13 models
│   ├── validation_framework_expanded18.csv      # Train/test split — expanded 18 models
│   └── validation_predictions.csv               # Actual vs. predicted Q-scores
└── framework/
    └── dimension_correlations.csv               # Pairwise correlations between Q-score dimensions

File Descriptions

`mechanistic/mechanistic_dimensions_all_models.csv`

The primary mechanistic profiling table. Contains all three observable dimensions extracted from reasoning chains for 19 models across 4 datasets.

Column	Description
`model_name`	Canonical model name
`params_b`	Parameter count in billions
`model_family`	Model family (DeepSeek, Qwen, GLM, etc.)
`training_method`	Training methodology (Distillation, RL, MCTS, etc.)
`cohort`	`original_13` or `expanded_18`
`dataset`	Benchmark (`gsm8k`, `boolq`, `arc-easy`, `gpqa`)
`diversity_delta_pct`	% change in vocabulary type-token ratio (300→max tokens)
`repetition_delta_pct`	% change in bigram repetition rate (300→max tokens)
`self_regulation`	Categorical: `Strong` / `Moderate` / `Weak`
`token_budget_utilization_pct`	% of allocated token budget consumed at max limit
`tokens_per_step`	Mean tokens per reasoning step
`diversity_score`	Normalized [0,1] diversity maintenance score
`regulation_score`	Normalized [0,1] self-regulation score
`decomposition_score`	Normalized [0,1] step decomposition efficiency
`quality_control`	Composite Quality Control dimension (0.5×diversity + 0.5×regulation)
`q_score`	Final Q-score ∈ [0,1] (0.7×quality_control + 0.3×decomposition)
`tier_number`	Quality tier 1–5
`tier_label`	`Excellent` / `Good` / `Moderate` / `Poor` / `Insufficient`

`qscores/token_efficiency_all_models.csv`

Summary table linking Q-scores to token efficiency (η) for all 19 models on GSM8K.

Column	Description
`q_score`	Training quality score
`peak_accuracy_gsm8k`	Best GSM8K accuracy achieved across all token limits
`peak_token_limit`	Token limit at which peak accuracy was achieved
`peak_avg_tokens`	Average tokens generated at peak condition
`efficiency_eta`	η = peak_accuracy / peak_avg_tokens

`qscores/qscore_cross_dataset.csv`

Q-scores computed per model per dataset. Use for cross-dataset generalization analysis (Table S8 in the paper). Same 19 models, 4 datasets.

`validation/validation_framework_original13.csv`

Full mechanistic features + Q-scores for the original 13 models, with split column indicating train (n=9) or test (n=4) assignment.

`validation/validation_framework_expanded18.csv`

Same as above but for the expanded 18-model cohort (original 13 + 5 new families: Microsoft, PRIME-RL, Qwen3-8B-Thinking, MiniMax×2).

`validation/validation_predictions.csv`

Framework validation results: actual Q-scores vs. predicted Q-scores from the held-out test sets. Covers both original (n=4) and expanded (n=10) holdout evaluations.

Column	Description
`q_score_actual`	Ground-truth Q-score
`q_score_predicted`	Q-score predicted by the framework
`q_score_abs_error`	\|actual − predicted\|
`tier_correct`	Whether tier assignment was correct
`evaluation_set`	`original_13_holdout` or `expanded_18_holdout`

`mechanistic/ablation_training_variants.csv`

Four Qwen2.5-7B training ablation variants studied in the paper: Base, Base+AnswerOnly, Base+LengthPenalty, Base+TerminationCalibration.

`framework/dimension_correlations.csv`

Pairwise Pearson correlations between the three Q-score base dimensions (Diversity, Regulation, Decomposition) on the training set (n=9 models). Justifies consolidating Diversity+Regulation into a single Quality Control dimension.

Models Covered

Model	Params	Cohort	Q-Score (GSM8K)	Tier
Skywork-o1-8B	8B	original_13	0.979	Excellent
Microsoft-Reasoning-4	14B	expanded_18	0.962	Excellent
QwenQwQ-32B	32B	original_13	0.944	Excellent
DS-R1-70B	70B	original_13	0.906	Excellent
PRIME-RL-Eurus-2-7B	7B	expanded_18	0.896	Excellent
DS-R1-32B	32B	expanded_18	0.868	Good
DS-R1-685B	685B	original_13	0.771	Moderate
Qwen3-32B-O1	32B	original_13	0.764	Moderate
DS-R1-14B	14B	original_13	0.755	Moderate
MiniMax-M2.5-229B	229B	expanded_18	0.656	Moderate
Nemotron-7B	7B	original_13	0.653	Moderate
GLM-4.5	355B (32B active)	original_13	0.610	Moderate
Qwen3-8B-Thinking	8B	expanded_18	0.561	Poor
GLM-4.7-355B	355B (32B active)	original_13	0.550	Poor
Kimi2.5-1T	1T (32B active)	original_13	0.540	Poor
DS-R1-7B	7B	original_13	0.512	Poor
MiniMax-2.1-229B	229B	expanded_18	0.465	Poor
Marco-o1-7B	7B	original_13	0.394	Poor
DS-R1-1.5B	1.5B	original_13	0.300	Insufficient

Q-Score Formula

Q = 0.7 × Quality_Control + 0.3 × Decomposition_Efficiency

where:
  Quality_Control     = 0.5 × diversity_score + 0.5 × regulation_score
  diversity_score     = 1 − |diversity_delta_pct| / 100
  regulation_score    = 1 − token_budget_utilization_pct / 100
  decomposition_score = 1 − norm(tokens_per_step)   [min-max, training set]

All three base dimensions are normalized to [0, 1] using min-max scaling on the training set (n=9 models). Test set normalization uses training-set bounds.

Key Results Reproducible from This Dataset

Paper Claim	File	Column
Q-score vs. efficiency r = +0.787	`token_efficiency_all_models.csv`	`q_score`, `efficiency_eta`
18.8× efficiency gap (Skywork vs. Marco)	`token_efficiency_all_models.csv`	rows 0 and 17
100% tier prediction accuracy (n=10 holdout)	`validation_predictions.csv`	`tier_correct`
GSM8K–GPQA r = −0.027	`qscore_cross_dataset.csv`	`dataset=gsm8k` vs `dataset=gpqa`
Diversity–Regulation r = +0.929	`dimension_correlations.csv`	all rows
DeepSeek distillation Q gradient	`mechanistic_dimensions_all_models.csv`	family=DeepSeek, dataset=gsm8k

Downloads last month: 4

Total file size:

36.7 kB