Recovered one transient empty-content API failure (JA26P1M09, ~300s
timeout) via targeted rerun + merge; rescored 318 -> 322.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
docs: add HTML leaderboard spec + implementation plan
Recovered 10 questions that originally hit the 32k max_tokens cap
(empty content / scored as API failures). Re-ran them at a 64k cap;
all 10 parsed (7 correct, 3 wrong), lifting the score from a false
316 to a genuine 339/360 — tied with qwen3.7-plus, 2 marks behind
Gemini 3.1 Pro.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
feat: add --provider-only flag to pin OpenRouter routing
Restricts OpenRouter routing to a specific provider (e.g. 'moonshotai'
for the official Kimi host) with fallbacks disabled, so a request fails
and is retried rather than silently routing to a third-party reseller
that may serve a different quantization.
Threaded through config["provider_only"] to both the initial call and
the re-prompt call. Adds two TDD tests for the request-body wiring.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
data: add Gemini 3.1 Pro JEE Advanced 2026 results (341/360, new #1)
Full 102-question paper. Tops the JEE Advanced 2026 board, edging
qwen3.7-plus (339). Run cost $10.97.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
chore: move NEET 2026 chart generator into gitignored charts/
The ad-hoc NEET 2026 chart script now lives in charts/neet_2026/
alongside its PNG output (uncommitted), matching the JEE Advanced 2026
convention. Remove the old tracked copy from scripts/ and update the
.gitignore comment to point at the new location.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add JEE Advanced 2026 runs for qwen/qwen3.7-plus (339/360) and
xiaomi/mimo-v2.5 (283/360), and rescore the existing gemma-4-31b (292)
and minimax-m3 (319) runs under the numeric-equivalence scoring fix
(+8 each from the '9.8'=='9.80' and '0.5'=='0.50' INTEGER questions).
Also registers the newly-run models (and other candidates) in
benchmark_config.yaml.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
fix: compare INTEGER answers by value, not string
Numerical answers (INTEGER/INTEGER_2) were scored with exact string
equality, so equal-but-differently-formatted values were marked wrong
(e.g. '9.8' != '9.80', '5' != '5.00'). Add integer_answers_match() which
falls back to a float comparison with a tight tolerance, used in both the
JEE Main and JEE Advanced INTEGER scoring paths. Range answers are
unaffected (they already use is_within_range/float).
Adds 4 tests covering the trailing-zero cases plus a false-positive guard
(9.79 != 9.80).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
docs: refresh README run examples and make them copy-friendly
Replace stale example model (openai/o3) and year (2024) with current
ones, put each command in its own code block for easy copying, and add
more usage examples: question-ID filter, --num_runs variance, temperature
override, a JEE Advanced run, and an Analyse results section covering the
leaderboard and aggregate_runs.py.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
docs: trim README into a lean dataset card (421 -> 146 lines)
Condense the dataset card to the essentials: description, composition
table, quick-start (load + run + leaderboard), scoring tables, merged
fields/answer-format section, contamination notes, citation, license.
Remove duplicated/generic content already covered in CLAUDE.md: the
6-step manual-usage walkthrough, Advanced Features prose, repo-layout
section, duplicate field list, and the generic troubleshooting section.
No dataset-specific information lost.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Replace the Current Data bullet list with a table (exam, year, set,
subjects, question count) plus a note on per-subject balance. Correct
NEET 2025 subjects to Biology to match the dataset labels (was listed
as Botany/Zoology).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Correct 17 numerical (INTEGER/INTEGER_2) answer keys for JEE Advanced 2026
against the official final answer keys (Paper 1 & 2). These were stored as a
flat two-element list (e.g. ["3.9", "4.1"]), which the scorer interpreted as
two exact-match alternatives, so any in-range value was marked incorrect.
Convert them to the nested range format ([["3.9", "4.1"]]) the evaluator
expects, matching the 2025 entries. Four ranges were also widened to match the
final key (provisional values had been narrower):
JA26P1P11 0.63-0.68 -> 0.63-0.70
JA26P2C11 1.45-1.50 -> 1.45-1.52
JA26P2C12 33.00-33.33 -> 31.60-33.33
JA26P2P13 694-695 -> 690-710
All other MCQ and exact-value keys verified to match the final key.
Add re-scored result runs against the corrected key:
minimax/minimax-m3 311/360 (86.4%)
google/gemma-4-31b-it 284/360 (78.9%)
Both runs have 2 (gemma) / 4 (minimax) remaining questions that the models
fail to answer: each loops to the token cap and returns empty content ac…
print() was competing with tqdm's progress bar, causing some
question result lines to be overwritten and disappear.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
fix: add MCQ_MATCHING and INTEGER_2 to prompt templates
Both new types were falling through to DEFAULT, giving the model
vague "Answer based on the question's format" instructions instead
of the correct single-option / numerical-value prompts.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
chore: track results/ in git, remove dead comment
1. MCQ_MULTIPLE_CORRECT negative penalty: -2 → -1 (actual paper says -1)
2. Add MCQ_MATCHING type for Section 4 Matching List questions (P1 Q13-16):
scores +4/-1/0 instead of the +3/-1/0 of plain MCQ_SINGLE_CORRECT
3. Add INTEGER_2 type for Section 4 paragraph-paired integers (P2 Q15-18):
scores +2/0 instead of +4/0
Metadata updated for all three years (2024/2025/2026):
- 30 questions retyped to MCQ_MATCHING
- 27 questions retyped to INTEGER_2
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- 102 images across Paper 1 (48 Qs) and Paper 2 (54 Qs)
- Fixed paragraph-question pairs: P2 Q15-18 for all three subjects
were duplicate images; split each into paragraph + individual question
- Fixed Chemistry P2 Q8 which had the wrong (Math) image in source repo
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Route all logging through tqdm.write() via _TqdmLoggingHandler so log
lines print above the bar rather than scrolling it off-screen — the
two-line display (bar + score) stays pinned at the bottom throughout
the run, matching the feel of apt/pip installs.
Add live cost accumulation to the score line: each completed question
adds its API cost, shown as $0.0042 (4dp below $0.10, 2dp above).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Shorten model name in desc (strip provider prefix, truncate at 26 chars)
- Add a second tqdm line below the bar showing running correct/total (%),
with ✓ / ~ / ✗ / skip counts updated after every question
- Drop the cryptic P!/A!/S shorthand from the postfix; failures only
appear in the score line and only when non-zero
- Unify counter variable names across initial and retry passes
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
perf: filter records before Dataset.from_list to avoid decoding images
Previously, exam/year/question_id filters ran via dataset.filter() after
the HF Dataset was constructed with decode=True, causing all 578 images
to be decoded to PIL just to evaluate metadata predicates. Filtering the
plain dict records list first means images are only decoded when actually
needed during inference.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
deps: upgrade huggingface-hub to 1.15.0 (adds Xet storage support)
Widen figure, extend x-limits, and push Flash Lite / Kimi K2.5 labels
down to clear the top-right cluster.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Show explicit dollar-formatted ticks instead of a single 10^0 mark,
include each model's cost next to its label, and place labels with
per-model offsets so the upper-right cluster (Gemini 3 Flash, GPT-5.5,
Grok 4.3, Gemini 3.1 Pro) no longer overlaps. Pareto-frontier points
get a darker outline and a soft green fill above the curve.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
scripts: add Grok 4.3 display name and color to chart generator
Reads live from results/*NEET_2026* and emits leaderboard, cost-vs-accuracy,
and per-subject heatmap PNGs into charts/neet_2026/. Charts directory is
gitignored — regenerate before publishing.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
