heihei/prune-to-prosper-data

Prune to Prosper - Embedding Dimension Analysis Data

This dataset contains experimental data for the paper "Dimensions Are Interchangeable: Evidence That Task-Aware Embedding Pruning Does Not Outperform Random Selection".

Dataset Structure

analyze/ — Per-Model Chunk Importance Analysis

Chunk-level importance scores for 3 models evaluated with win_size=2 (512 chunks for 1024-dim models):

File	Model	Size
gte-large-en-v1.5.json	GTE-Large	4.7 MB
stella_en_400M_v5.json	Stella EN 400M	4.7 MB
roberta-large-InBedder.json	Roberta-Large-InBedder	4.7 MB

Each file contains per-task chunk importance scores, including:

• task_name → task → split_win_size → win_size → chunk_result (512 scores)
• defult_score, random_score, sort_score at task level

task_similar/ — Cross-Task Dimension Ranking Transfer

Dimension ranking transfer data for 12 models, showing retention when using task A's ranking to prune for task B.

Each JSON file contains task pairs with:

• Source task dimension ranking
• Target task retention ratio
• Spearman rank correlation between rankings

mteb/ — MTEB Evaluation Results

Full MTEB benchmark results for 13 embedding models:

• gte-large-en-v1.5/, stella_en_400M_v5/, roberta-large-InBedder/ (detailed models)
• bge-m3/, gte-base/, gtr-t5-large/, instructor-large/ (additional models)
• mxbai-embed-large-v1/, Qwen3-Embedding-0.6B/ (recent models)
• roberta-large/, bart-base/ (non-contrastive models)
• gte-Qwen2-1.5B-instruct/, jina-embeddings-v3/ (instruction-tuned models)

experiment_results/ — Analysis Outputs

Key experimental analysis results:

File	Description
analysis_results.json (906K)	Main chunk analysis results
near_optimal_mask_analysis.json (1.7M)	Near-optimal mask degeneracy analysis
universal_mask_analysis.json	Universal mask transfer experiment
basis_sensitivity_gte-large.json	Basis independence for GTE-Large
basis_sensitivity_stella.json	Basis independence for Stella
magnitude_analysis.json	Magnitude pruning analysis
reviewer_response_analysis.json	Reviewer response experiments
all_methods_comparison.json	All 5 methods comparison

Key Findings (from this data)

1. Normalized entropy = 0.988–0.993: Dimension importance is nearly uniform
2. Optimized-Random gap = +2.2–5.0%: Task-aware pruning barely helps
3. Cross-task retention = 95–100%: Despite ρ ≈ 0.001 ranking correlation
4. Basis independence: Sequential-Random gap < 1% under all tested rotations
5. 31.6% of random masks within 1% of oracle: Near-optimal mask degeneracy

Usage

import json

# Load chunk importance for GTE-Large
with open("analyze/gte-large-en-v1.5.json") as f:
    data = json.load(f)

# Get chunk importance for a specific task
task = "Banking77Classification"
scores = data["task_name"][task]["split_win_size"]["2"]["chunk_result"]
print(f"Number of chunks: {len(scores)}")
print(f"Top-10 most important chunks: {sorted(range(len(scores)), key=lambda i: scores[i], reverse=True)[:10]}")

Related

• Paper: GitHub - ngyygm/prune-to-prosper
• MTEB Benchmark: https://github.com/embeddings-benchmark/mteb