Skip to content
System One

chunklaya

Measures how Laya's decision head stops attending to input past roughly 200 tokens, then fixes it by chunking documents into passages and screening each one. Tested on 400 documents, and against Jev on a public needle benchmark.

myxamediyar/chunklaya on GitHub

Stars
3
Forks
0
Language
Python
License
Apache-2.0
Last push
23 Sep 2026

Read from GitHub on . Counts change daily.

How it uses Jev

ap.add_argument("--n", type=int, default=0, help="limit items per size (0 = all)")
ap.add_argument("--model", default="jev-latest")
ap.add_argument("--out", default=str(ROOT / "results" / "jev-ours"))
ap.add_argument("--url", default="https://api.typesafe.ai/v1/systemone")
ap.add_argument("--retries", type=int, default=5)
ap.add_argument("--dry-run", action="store_true", help="price the run, send nothing")
args = ap.parse_args()


def api_key():
    k = os.environ.get("TYPESAFE_API_KEY")
    if k:
        return k.strip()
    p = Path.home() / ".garlic" / "typesafe-api-key"
    if p.exists():

View in eval/run_jev_needle.py

Open on GitHub

More like this

Use cases this is tagged with