sample-evaluation-scripts/README.md

# Evaluation scripts

Three scripts for scoring model responses against the golden answers. Two are
deterministic and need no model, no API key and no network; the third uses an
LLM judge.

| Script | Protocol | Needs a model? |
|---|---|---|
| `exact_substring.py` | Golden answer must appear as one contiguous substring | No |
| `word_overlap.py` | Every golden word must appear, in any order | No |
| `llm_judge.py` | A judge model decides semantic equivalence | Yes |

Only `pandas` is required for the deterministic scripts.

## Input format

Each script takes one CSV: a language file from this dataset with a `response`
column added.

| column | source |
|---|---|
| `question` | dataset |
| `answer` | dataset (golden answer) |
| `Domain` | dataset |
| `response` | your model's raw output |

```bash
python exact_substring.py --responses my_model_english.csv
python word_overlap.py    --responses my_model_english.csv --out scored.csv
python llm_judge.py       --responses my_model_english.csv --limit 50
```

Each prints per-domain and combined accuracy. Combined is the micro-average
over all pooled questions, which is identical to weighting each domain by its
size. `--out` writes per-question verdicts.

`llm_judge.py` ships with `call_model()` as a stub — implement it for your
backend (the docstring sketches a local `transformers` pipeline and an
OpenAI-compatible endpoint) and use temperature 0, or scores will not
reproduce.

---

## 1. Exact substring

The entire logic:

```python
is_correct = answer.lower() in response.lower()
```

The golden answer must appear as **one contiguous run of characters**.
Lowercasing is the only normalisation — no trimming, no punctuation handling,
no tokenisation, no word boundaries.

| Golden | Response | Verdict | Why |
|---|---|---|---|
| `Ruru Jataka` | `The answer is the Ruru Jataka, depicted at Bharhut.` | correct | surrounding prose is irrelevant |
| `ruru jataka` | `RURU JATAKA` | correct | case-insensitive |
| `गंगा` | `गंगा नदी` | correct | works for Devanagari |
| `Ruru Jataka` | `Jataka Ruru` | wrong | order matters |
| `Narmada valley` | `the Narmada river valley` | wrong | must be contiguous |
| `Delhi ` (trailing space) | `Delhi` | wrong | golden answer is not stripped |
| `amalak` | `Amalaka` | correct | matches inside a longer word |
| `No` | `There is **no** such temple` | correct | false positive: no word boundary |
| `Delhi` | `Delhi is not the answer; it's Mumbai` | correct | false positive: mention is not assertion |

**What it measures:** whether the model reproduced the golden phrase verbatim,
including word order and internal spacing. A phrase-fidelity test.

**Failure modes.** False negatives dominate and are mostly cosmetic —
reordering, an inserted qualifier, stray whitespace in the golden answer — so
the score is a lower bound. False positives are rarer but more damaging:
nothing anchors the match to a word boundary or to what the model actually
asserted, so short golden answers can match inside unrelated words, and a
response that names the golden answer only to reject it still scores correct.

---

## 2. Word overlap

Both sides are lowercased, split on whitespace, stripped of leading and
trailing punctuation (`delhi.` → `delhi`), and compared as **sets**:

```python
is_correct = set(golden_words) <= set(response_words)
```

All golden words must appear, in any order, anywhere in the response. Extra
words are free — a subset test, not equality. The output CSV also carries
`matching_words` and `total_golden_words`, giving partial credit that the
binary verdict hides.

| Golden | Response | Verdict | Count | Why |
|---|---|---|---|---|
| `Narmada valley` | `valley Narmada` | correct | 2/2 | order is irrelevant |
| `Narmada valley` | `the Narmada river valley in India` | correct | 2/2 | insertions are free |
| `Narmada valley` | `Narmada` | wrong | 1/2 | every golden word required |
| `1947 to 1947` | `it was 1947` | wrong | 1/2 | duplicates collapse to a set |
| `Amalaka` | `amalak` | wrong | 0/1 | whole-token match, no stemming |
| `Chola ideals` | `Chola and Hoysala ideals` | correct | 2/2 | false positive: different claim |

**What it measures:** whether the response contains the golden vocabulary,
disregarding order, position, and anything said between the words. A
content-word recall test.

**Failure modes.** Extra words are never penalised, so a verbose answer that
happens to include every golden word passes — this is the main false-positive
channel and it grows with response length. The comparison is a set rather than
a multiset, so repetition is never checked. There is no stemming, which matters
a great deal for Indic morphology.

---

## How the two differ

**Neither is a looser version of the other.** They disagree in both directions,
because they relax and tighten different axes.

| Axis | Exact substring | Word overlap |
|---|---|---|
| Word order | must match | irrelevant |
| Inserted words inside the phrase | fails | passes |
| Extra words elsewhere | passes | passes |
| Sub-word match (`amalak` / `Amalaka`) | passes | fails |
| Whitespace noise in golden answer | fails | tolerated |
| Unit of comparison | character run | whole token |
| Partial credit reported | no | yes |

### Real disagreements

From an actual run (Sarvam 30B on the Art domain, 233 rows — exact 59/233,
word overlap 56/233). The near-identical totals hide rows that flip in
*opposite* directions.

Exact passes, overlap fails — morphological variants where the golden string
sits inside a longer word:

| Golden | Response |
|---|---|
| `amalak` | `Amalaka` |
| `deul` | `Deula` |
| `mithun` | `Mithuna` |
| `Dipankar` | `Dipankara Buddha` |
| `scroll painting` | `Scroll paintings` |

Overlap passes, exact fails — all golden words present, but not contiguous:

| Golden | Response |
|---|---|
| `Narmada valley` | `Narmada River valley` |
| `Chola ideals` | `Chola and Hoysala ideals` |

Every one of the first group is arguably a correct answer that exact substring
catches and word overlap misses on a technicality. In the second group,
`Narmada River valley` is correct and `Chola and Hoysala ideals` is not — word
overlap gets one right and one wrong for the same reason.

### Reading the two scores together

Because the metrics are near-orthogonal, the pair is more informative than
either alone:

- **Both pass** → high confidence the answer is right.
- **Both fail** → high confidence it is wrong, or phrased very differently.
- **Exact only** → almost always an inflection difference; usually a correct
  answer under-counted by word overlap.
- **Overlap only** → the golden words are all there but rearranged. Could be a
  correct paraphrase or a genuinely different claim. This bucket needs human or
  judge review.

Treat both numbers as **lower bounds**. Neither understands paraphrase,
synonymy, negation, or numeric equivalence. For that, use `llm_judge.py`.

---

## Caveats for both deterministic scripts

1. **An empty golden answer scores correct** in both (`"" in x` is `True`, and
   an empty set is a subset of anything). Both scripts warn on stderr if the
   input contains one.
2. **Golden answers are not stripped**, so trailing whitespace breaks exact
   substring outright.
3. **No Unicode normalisation.** `.lower()` is a no-op for Indic scripts, and
   NFC versus NFD forms of the same word compare unequal. Tokenising is fine,
   but equality is fragile for Indic text — consider normalising to NFC before
   scoring if your responses come from mixed sources.
4. **Blank responses score incorrect**, and are counted separately in the
   output so that missing data is distinguishable from wrong answers.