raihan-js/flipgate-results

FlipGate Results

Per-item results of FlipGate, a release gate that counts right-to-wrong answer flips between a baseline and a candidate model. Qwen2.5-3B-Instruct in bf16, AWQ and GPTQ-Int4 (HF generate), plus an HF-generate-vs-vLLM engine control (bf16, three vLLM passes) and early llama.cpp rows. 49 runs, 14,368 items; greedy decoding. Every prompt, response and score is here, so every number in the repo can be recomputed.

Rebuilt 2026-10-05. The previous revision of this dataset (and its card) described a 30-item pilot. This revision is a flat re-export of everything, with the corrected GSM8K re-run added and the superseded runs flagged in the status column. The old parquet is in this repo's git history.

Tables

from datasets import load_dataset
items = load_dataset("raihan-js/flipgate-results", "items", split="train").to_pandas()
runs = load_dataset("raihan-js/flipgate-results", "runs", split="train").to_pandas()
cur = runs[runs.status.str.startswith("current")]

• items: run_id, model, engine, dataset, batch_size, item_id, prompt, response, score, finish_reason, n_new_tokens, score_v1, timestamp. finish_reason (stop or length) and n_new_tokens exist only for the runs made with the corrected harness.
• runs: run_id, model, engine, engine_version, dataset, batch_size, num_items, accuracy, accuracy_v1_scorer, max_new_tokens, truncated_rate, manifest_version, created, status.

Which runs to use (status)

status	runs	what it means
current: 1,024-token cap, finish reasons recorded	Qwen2.5-3B-{bf16,AWQ,GPTQ-Int4}_gsm8k_v2, Qwen2.5-3B-bf16_gsm8k_v2_bs8	the corrected GSM8K sweeps (1,000 items at batch size 1) and the batch-size-8 noise-floor run (200 items)
current	IFEval (541 prompts), FedProc registry check (155), BFCL (550) runs	short outputs; finish reasons were not recorded
superseded: 256-token cap, most answers truncated	the three original 1,000-item GSM8K runs	kept so the truncation can be re-checked; do not compare with the v2 runs
withdrawn: ...	the two llama.cpp 200-item runs	256-token cap and a different system prompt from the HF runs
early small run	pilot GSM8K runs (30-200 items)	generation cap not recorded

Headline results (corrected GSM8K, 1,000 items)

Comparison	Accuracy	Right→Wrong	Wrong→Right	McNemar p
bf16 vs AWQ	79.7% → 76.2%	91	56	0.0050
bf16 vs GPTQ-Int4	79.7% → 76.0%	91	54	0.0028

All 3,000 responses ended naturally. Batch-size noise floor (bf16, batch 1 vs 8, 200 items): 124 of 200 responses differ in text, 6 right→wrong and 7 wrong→right (3.0% [1.0%, 5.5%]). The original 256-token runs reported bf16 33.0% / AWQ 43.1% / GPTQ 37.3% and are flagged superseded. The strict (score_v1) and robust (score) answer extractors disagree on the sign of AWQ's accuracy change; see the repo README.

IFEval, FedProc and BFCL results are in the repo README (FedProc: AWQ +34 new hallucinated clause numbers, p = 0.001).

Limits

One model family and size; greedy decoding; HF generate for the quantised models; the noise floor is from 200 items on one stack.

Citation

@software{flipgate2026,
  author = {Raihan Sikder},
  title = {FlipGate: a release gate for quantised LLMs},
  year = {2026},
  url = {https://github.com/raihan-js/flipgate}
}