ggranberry/lemmanaid-afp-reruns
Lemmanaid AFP-pool Reproducibility Reruns Reproducibility study for claude-opus-4-5 on the yalhessi/lemexp-commerical-llm-experiment benchmark, using an AFP demo pool (honest eval — no train/test theory leakage). Companion to ggranberry/lemmanaid-commercial-results, which holds the earlier shot-count + retrieval sweeps under test-LOO. Configs Two configs, one per benchmark domain: Config Source HF config Test rows octonions template_octonions_2026… See the full description on the dataset page: https://huggingface.co/datasets/ggranberry/lemmanaid-afp-reruns.
Lemmanaid AFP-pool Reproducibility Reruns
Reproducibility study for claude-opus-4-5 on the `yalhessi/lemexp-commerical-llm-experiment` benchmark, using an AFP demo pool (honest eval — no train/test theory leakage).
Companion to `ggranberry/lemmanaid-commercial-results`, which holds the earlier shot-count + retrieval sweeps under test-LOO.
Configs
Two configs, one per benchmark domain:
Load with:
from datasets import load_dataset
ds = load_dataset("ggranberry/lemmanaid-afp-reruns", "octonions", split="template_t1_r1")Splits
Each config has 10 splits = 2 tasks × 5 runs.
All runs share identical prompts (5-shot nearest-by-symbols retrieval, seed=0, demos drawn from yalhessi/lemexp-task1-v3/{template_afp,lemma_object_afp} train split). Only API sampling differs.
Row schema
{"custom_id": "...", "match": bool, "prediction": "...", "gold_output": "...", "lemma_name": "..."}Raw results
For each split there's also a results_ext_*.jsonl file at repo root (not mounted as a config) containing the unscored predictions plus stop_reason and token usage metadata.
Results
5-shot nearest, AFP-pool, opus-4-5, seed=0:
- Sampling variance is small: std across 4 T=1 reruns ≤1.2pp; T=0 falls inside the T=1 band.
- pass@4 headroom: gap pass@4 − pass@1 ranges 0.9–4.3pp — recoverable with a verifier/reranker.
- Note on pass@k: 4 independent T=1 samples per query; "pass" = any match.
Method
- 5-shot prompt; demos picked by Jaccard symbol-overlap from AFP train pool (no test theory overlap)
- Same
seed=0across all 5 runs → identical demos, identical prompts; only API sampling differs - Anthropic Message Batch API (50% discount, ≤24h SLA)
- Exact-match scoring against gold
Pipeline source: https://github.com/... (private)
