Upload README.md with huggingface_hub

README.md

@@ -1,3 +1,4 @@
+
 ---
 dataset_name: cleaned-plotqa-v2-difficulty
 tags:
@@ -6,59 +7,32 @@ tags:
 - curriculum-learning
 - difficulty-estimation
 - rule-based
-dataset_info:
-  features:
-  - name: image
-    dtype: image
-  - name: user
-    dtype: string
-  - name: assistant
-    dtype: string
-  - name: difficulty_tier
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 7194085718.375
-    num_examples: 199293
-  download_size: 121124922
-  dataset_size: 7194085718.375
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
 ---
 
-# Cleaned-PlotQA v2 with difficulty tiers (balanced, rule-based)
+# Cleaned-PlotQA v2 with difficulty tiers (calibrated, rule-based)
 
-This repository augments jrc/cleaned-plotqa-v2 with one additional column, difficulty_tier ∈ {easy, medium, hard}. The classifier uses strong PlotQA-oriented rules emphasizing numeric reasoning, extremum/trend/slope semantics, visual grounding (axes/legend/lines/bars), units/notation, and multi-entity disambiguation. Tiers are balanced using quantiles computed over the full dataset for robust stage sizing in curriculum learning.
+This repository augments jrc/cleaned-plotqa-v2 by adding a single column difficulty_tier ∈ {easy, medium, hard}, tuned for numeric and visual reasoning in scientific plots and calibrated on a 1,000-example sample with a robust tie-break fallback to prevent score collapse.
 
 ## Tier counts
-
 - easy: 199293
 - medium: 0
 - hard: 0
 - total labeled: 199293
 
 ## Criteria summary
+This release adds a single column difficulty_tier ∈ {easy, medium, hard} using a robust, PlotQA-oriented scoring with a rank-based fallback and balanced cutoffs:
 
-This release adds a single column difficulty_tier ∈ {easy, medium, hard} using a deterministic, PlotQA-oriented scoring function with balanced thresholds:
+Calibration (on 1,000-sample):
+- Primary score weights numeric operations (sum/diff/ratio/percent/average), extremum/trend/slope cues (max/min/peak, slope/rate-of-change), visual grounding (axes/legend/lines/bars), units and formats (%, scientific notation, ranges), and multi-entity hints (both/all/each/every/combined/grouped/stacked/multi, and/vs/per/between) [PlotQA context].
+- If the primary score distribution collapses (near-constant), a deterministic rank-based fallback (counts of ops/extremum/units/percent/decimal/scientific/range/multi-entity/position/visual/color/length) is normalized and blended to ensure separation.
+- Balanced thresholds are set from the sample at 40th and 80th percentiles:
+  easy: score ≤ 0.000000
+  medium: 0.000000 < score ≤ 0.000000
+  hard: score > 0.000000
 
-Scoring highlights:
-- Numeric reasoning: detection of arithmetic/aggregation (sum, difference, ratio, percentage, average), stacked operations, and comparison/threshold cues.
-- Extremum/trend/slope: emphasis on max/min/peak/valley and slope/rate-of-change; any “line/curve + extremum” is never classified as easy.
-- Visual grounding: references to axes, ticks, legend, lines/bars/curves, and positional terms (left/right/top/bottom/adjacent) raise difficulty.
-- Units and formats: recognition of %, scientific notation, unit tokens (k/M/B, °C, km, kg, Hz, etc.), ranges (10–20), and uncertainty (±, std dev, variance).
-- Series disambiguation: color terms and multi-entity words (both/all/each/every/together/combined/grouped/stacked/multi) increase difficulty.
-
-Balancing:
-- Tiers are assigned by data-driven quantiles over the entire dataset:
-  - easy: score ≤ 0.0000 (≈ bottom 40%)
-  - medium: 0.0000 < score ≤ 0.0000 (≈ middle 40%)
-  - hard: score > 0.0000 (≈ top 20%)
-This ensures a balanced distribution even if raw scores are skewed.
+Assignment across the full dataset:
+- The same blended scoring and fixed cutoffs are applied to every example to produce tiers suitable for curriculum training.
 
 ## Notes
-
 - Only one new column is introduced; all original fields remain unchanged.
-- The balancing by quantiles is designed to produce sensible stage sizes across easy/medium/hard and can be retuned (e.g., 30/40/30) by adjusting the quantile cut points.
+- The calibration procedure is designed to yield balanced tiers even when questions are short or templated, which otherwise causes naive rule-based scoring to concentrate mass in a single class.