Datasets & Ground Truth

The most common reason an eval lies to you is a bad dataset. Metrics get all the attention, but the dataset is where credibility is quietly won or lost.

What makes an eval set trustworthy

• Representative — mirrors the real input distribution (queries, formats, languages). An eval of clean, easy inputs gives a flattering, useless number.
• Sufficiently sized — big enough that the aggregate is stable run-to-run (Chapter 1's lesson).
• Sliced — tagged by segment so you see where quality lives, not just an average that hides failures.
• Hard where it matters — deliberately includes the failure-prone, high-stakes cases. The point is to catch problems, not feel good.

Goldens rot; contamination inflates; synthetic helps with care

A golden dataset is a curated, human-blessed yardstick — and it decays as the product and users change. Treat it as a maintained, versioned, owned asset. Watch for data contamination: if your eval cases were in the model's training data (common with public benchmarks), the score is inflated and meaningless for your use case — prefer private, recent, or custom sets. Synthetic data (an LLM generating plausible cases) is great for bootstrapping breadth, but skews toward what the model finds natural; use it to scale coverage, and real human-curated data to anchor truth.

Lab code — runs free

from statistics import mean

# Per-case scores, already tagged by slice
SCORED = [
    {"slice": "en",   "score": 0.95}, {"slice": "en",   "score": 0.92},
    {"slice": "en",   "score": 0.97}, {"slice": "en",   "score": 0.90},
    {"slice": "es",   "score": 0.61}, {"slice": "es",   "score": 0.55},
    {"slice": "code", "score": 0.40}, {"slice": "code", "score": 0.48},
]

overall  = mean(r["score"] for r in SCORED)
by_slice = {s: mean(r["score"] for r in SCORED if r["slice"] == s)
            for s in {r["slice"] for r in SCORED}}

print(f"overall: {overall:.2f}")
for s, v in sorted(by_slice.items(), key=lambda kv: kv[1]):
    flag = "   <-- WEAK" if v < 0.70 else ""
    print(f"  {s:5} {v:.2f}{flag}")

overall: 0.72
  code  0.44   <-- WEAK
  es    0.58   <-- WEAK
  en    0.94