ML
MLP Project · Kaggle Analytics
2026 T1 · L1 Viva Eligibility

Competition Score Analytics
& Grading Intelligence

F1-Macro scores · 2,339 participants · Outlier-adjusted grading formula · IITM BS Data Science 2026

👥Total Registered
2,339
All submissions in system
100%
Viva Eligible
1,399
Score ≥ 0.800 · Eligible Yes
59.8%
📉Below Cutoff
906
score < cutoff label
38.7%
⚠️Reg. Failed
34
Registration unsuccessful
1.5%
μMean Score
0.8161
Eligible participants only
σ = 0.0106
~Median Score
0.8145
P50 of eligible cohort
IQR 0.0161
Top Score
0.8711
saanvitiwari · z=5.18
Outlier
P95P95 Score
0.8345
Proposed adjusted top
Fair Cap
Score Distribution — Eligible Participants
F1-Macro score histogram (n=1,399 · above 0.800 cutoff)
Participation Breakdown
By eligibility category
59.8%
ELIGIBLE
Eligible (1,399)
Below Cutoff (906)
Reg. Failed (34)
Score Percentile Profile
Distribution statistics for eligible cohort
Min (cutoff floor) — 0.80000.8000
P25 (25th percentile)0.8074
P50 / Median0.8145
Mean (μ)0.8161
P75 (75th percentile)0.8235
P900.8311
P95 → Proposed Fair Cap0.8345
Max (raw top score)0.8711 ⚠
Box-Whisker Plot · Eligible Scores
Visual spread with outlier annotation
⚠️
Outlier Detected · z-score = 5.18σ
The top score of 0.8711 sits 5.18 standard deviations above the mean — statistically extreme. The next score (0.8646) is 3σ out. Using this as the grading ceiling compresses 99.8% of participants into the lower half of the scale, which is inequitable. Recommendation: cap at P95 = 0.8345.
Grade Distribution Comparison
Original (top=0.87) vs Adjusted (top=P95=0.8345)
Original (top=0.87)
Adjusted (top=P95)
Mean Grade Shift
Impact of top-score cap on average grade
ORIGINAL FORMULA (top = 0.87)
22.6%
Average grade
Most students score <30%
Extremely compressed scale
ADJUSTED FORMULA (top = P95 = 0.8345)
45.8%
Average grade
Balanced distribution
Reflects true performance
⚠ Original
Top0.8700
Mean Grade22.6%
>50% grade~2%
✓ Adjusted
Top0.8345
Mean Grade45.8%
>50% grade~38%
Top 10 Leaderboard
Showing both original and adjusted grades side by side
# Kaggle Username F1-Macro Score Grade (Original 0.87) Grade (Adjusted P95) Δ Impact
📐 Grading Formula Analysis
Fair Grading Formula Design
Outlier-robust score normalization with statistical justification

The base formula normalizes participant scores relative to a top-scorer benchmark, rewarding performance above the minimum cutoff. The challenge: when the top score is a statistical outlier, the formula systematically under-rewards all other participants.

BASE FORMULA
grade = (score − 0.800) / (top_score − 0.800)
∀ score ≥ 0.800  |  scores < 0.800 are excluded from this formula
ADJUSTED FORMULA (RECOMMENDED)
grade = min(score, P95) − 0.800) / (P95 − 0.800)
where P95 = 0.8345  |  scores above P95 are capped at grade = 1.0 (full marks)
❌ Problem with top=0.87
  • Top score is 5.18σ above mean — extreme outlier
  • Only 3 of 1,399 scored above 0.86 (0.2%)
  • 99.8% of cohort is bunched below 50% grade
  • Mean grade = 22.6% — most students at ~D level
  • Does not reflect actual competition difficulty
  • Single exceptional result penalizes everyone else
✅ Benefits of P95 cap
  • P95 is statistically robust — not swayed by outliers
  • Top 5% of students still get full credit (grade=1.0)
  • Mean grade rises from 22.6% → 45.8%
  • Score spread better reflects cohort's actual range
  • Consistent with academic norm-referencing practices
  • Denominator shrinks: 0.8345−0.800 = 0.0345 vs 0.07
P95 = 0.8345
z(top) = 5.18σ
3 scores > 0.860
IQR = 0.0161
14 scores in 0.84–0.86 range
Δ mean grade = +23.2pp
📖 Data Story
The Story of 2,339 Competitors
A narrative through the numbers

When 2,339 students submitted their machine learning models to the Kaggle competition, the results painted a vivid picture of performance, determination, and an unexpected twist at the top. The 0.800 cutoff emerged as a natural watershed: it divided the cohort nearly 60/40 — 1,399 crossed the line, 906 didn't. The successful students clustered tightly between 0.800 and 0.840, reflecting a cohort that largely converged on similar modeling approaches. Then, high above the rest, one student (saanvitiwari) achieved 0.8711 — 5.18 standard deviations beyond the mean. This exceptional score, while impressive, created a statistical paradox: it made everyone else look average. The grading formula, sensitive to this single ceiling value, compressed 99% of legitimate effort into a narrow grade band. The fix is elegant: cap the denominator at P95. The top 5% still earn full marks. Everyone else gets grades that reflect their genuine standing — not their distance from a single outlier.

59.8%
Cleared the 0.800 barrier
1,399 of 2,339 participants met the minimum F1-Macro threshold required for L1 Viva eligibility. A meaningful majority, but not comfortable.
5.18σ
The outlier problem, quantified
The top score sits 5.18 standard deviations above the mean. In a normal distribution, this happens 1 in 10 million times. It's a genuine statistical anomaly.
+23pp
Grade fairness restored
Switching to the P95-capped formula raises average grades from 22.6% to 45.8% — adding 23 percentage points of fairness without changing who tops the board.
0.0106
A tight, competitive field
The standard deviation among eligible scorers is just 0.0106. The middle 50% spans only 0.016 points. This is an intensely compressed competition.
906
The ineligible cohort's story
Most ineligible students scored near 0.18 — suggesting a default/baseline prediction. 138 came agonizingly close (0.78–0.80) but fell just short of the cutoff.
~19
Atharv's rank (atharvkhare)
At 0.83955, atharvkhare ranked approximately 19th overall — solidly in the top 1.4% of all participants. With the adjusted formula, this earns a strong full grade.
Ineligible Score Distribution
906 below-cutoff + 34 registration failures
Key insight: A massive spike near 0.18 (475 students) likely represents a default/baseline model output. The cluster near 0.78–0.80 (213 students) represents students who were close but missed the threshold.
Score Band Heatmap
Score count by range across all participants