Hawky-ai/feather-db-benchmarks

Feather DB — Benchmark Result Audit Trail

Per-run JSON results for Feather DB v0.8.0 on canonical retrieval and memory benchmarks. Every number cited in the report and arXiv paper maps back to one of these files.

Why this dataset exists

Most "AI memory" tools publish marketing numbers without a reproducible audit trail. We disagree with that practice. Every JSON here is a complete record of one benchmark run — config, environment, per-axis scores, failure traces, wall time, hardware. If you re-run the same command and get different numbers, please open an issue with your JSON.

Headlines (from these files)

Run	Variant	Answerer	Overall	File
LongMemEval, decay on, GPT-4o	S	gpt-4o	0.693	longmemeval__s__20260426_110536.json
LongMemEval, decay on, Gemini-Flash	S	gemini-2.5-flash	0.657	longmemeval__s__20260426_025723.json
LongMemEval, decay on	oracle	gemini-2.5-flash	0.670	longmemeval__oracle__20260425_200250.json
LongMemEval, no decay	oracle	gemini-2.5-flash	0.656	longmemeval__oracle__20260425_191858.json
SIFT1M @ 500K, ef sweep	sift1m	n/a (ANN only)	recall@10=0.972	vector_ann_real__sift1m__20260425_154057.json

Schema (each JSON)

{
  "scenario": "longmemeval | vector_ann | vector_ann_real",
  "dataset":  "oracle | s | sift1m | siftsmall | synthetic",
  "n":        500,        // questions or vectors
  "dim":      1536,       // embedder / vector dimensionality
  "feather_version": "0.8.0",
  "python_version":  "3.12.x",
  "platform":        "Darwin arm64",
  "commit":          "<short git sha>",
  "started_at":      <unix timestamp>,
  "wall_seconds":    16326,
  "params":          { ... knobs (k, ef, decay, embedder, judge) ... },
  "metrics":         {
    "overall":               0.693,
    "by_axis":               { /* per-axis means */ },
    "by_question_type":      { /* per-type means */ },
    "per_q_seconds_p50":     32.1,
    "per_q_seconds_mean":    32.4,
    "n_questions":           500,
    "n_scored":              495,
    "n_failures":            5,
    "failures_sample":       [ /* up to 10 failure rows */ ],
    "embedder":              "azure_text-embedding-3-small_d1536",
    "judge":                 "llm_judge=gemini/gemini-2.0-flash_ans=azure/gpt-4o-feather",
    "decay_engaged":         true,
    "decay_half_life_days":  14.0,
    "decay_time_weight":     0.4
  }
}

How to load

from datasets import load_dataset

ds = load_dataset("Hawky-ai/feather-db-benchmarks")
print(ds)
# Each row corresponds to one benchmark run.
# Filter by scenario / dataset / answerer in your analysis.

Or just clone and parse directly:

import json, glob
runs = [json.load(open(p)) for p in glob.glob("*.json")]

How to reproduce

The harness lives at bench/ in the source repo. Each scenario has a single-line CLI:

# LongMemEval (the headline)
python -m bench run longmemeval --dataset s --limit 0 \
    --embedder openai \
    --answerer-provider azure --answerer-model gpt-4o-feather \
    --judge llm --judge-provider gemini --judge-model gemini-2.0-flash \
    --decay-half-life 14 --decay-time-weight 0.4 --k 10

# SIFT1M ANN sweep
python -m bench run vector_ann_real --dataset sift1m \
    --n 500000 --queries 1000 --k 10 \
    --ef-sweep "10,50,100,200"

Result JSONs auto-write into bench/results/ and a Markdown rolled-up table into bench/reports/latest.md. New JSONs go here.

What's NOT in here

• The LongMemEval source dataset itself (lives at xiaowu0162/longmemeval-cleaned, not redistributed by us).
• The SIFT1M source dataset (downloads from corpus-texmex.irisa.fr).
• API keys, embedder weights, LLM responses verbatim. We log enough metadata to reproduce; we don't republish closed-model outputs.

License

MIT. Same as Feather DB.

Citation

If you use these results in academic work, cite the Feather DB paper:

@article{featherdb2026,
  title   = {Feather DB: An Embedded Vector Database with Adaptive Temporal
             Decay and Hybrid BM25/Dense Retrieval},
  author  = {Hawky.ai},
  year    = {2026},
  journal = {arXiv preprint},
  url     = {https://github.com/feather-store/feather/blob/master/docs/featherdb_paper.pdf}
}

And cite the underlying benchmarks:

• LongMemEval: Xu et al., 2024 (arXiv:2410.10813)
• SIFT1M: Jégou, Douze, Schmid, 2011 (INRIA report)

---

Maintained by Hawky.ai. Last update: 2026-04-26.