eye-grep — log token-classification gold set

Token-level labels for eye-grep, a log colorizer that tags each content token of a server-log line so a renderer can highlight ids, timestamps, IPs and repeated strings. These are the sets used to train and evaluate the eye-grep taggers (opsbr/eye-grep-deberta-v3-small and the distilled opsbr/eye-grep-electra-small).

Fully synthetic and self-contained — generated by a deterministic template engine (loop/synth_gold.py), with no third-party log data. Every value is produced by a typed slot generator, so the gold labels are exact by construction (no LLM labeling, no alignment error).

Splits

split	rows	structures
train (gold_train.jsonl)	2,943	93 templates across 58 system families
validation (gold.jsonl)	146	18 held-out templates (disjoint structures → tests generalization)

System families span web servers (Nginx, Apache, HAProxy), databases (PostgreSQL, MySQL, Redis, MongoDB, Cassandra, ClickHouse), messaging (Kafka, RabbitMQ, NATS), container/orchestration (Kubernetes, Docker, Envoy, Istio, Traefik), big-data (Spark, Hadoop, HDFS, Flink), cloud (Lambda, CloudTrail), OS/syslog (Linux, Windows, macOS, firewall, DNS, mail), and application logs (JSON, logfmt, Java, Go, Python, Rails) — in formats from syslog and Apache-combined to JSON and key=value.

Format

JSON Lines — one object per line:

{"system": "Service", "line": "2026-03-16T22:58:29 INFO order-service request_id=tok_9fKd ...",
 "content_tokens": ["2026-03-16T22:58:29", "INFO", "order-service", ...],
 "gold": ["TIMESTAMP", "LEVEL", "WORD", ...]}

• system — synthetic source family (for diversity / the held-out split).
• line — the generated log line.
• content_tokens — content tokens of the line (separators dropped), per eye-grep's frozen tokenizer (train/spec.py).
• gold — one tag per content token, index-aligned to content_tokens.

Label schema (11 tags)

PUNCT WORD NUM RAND IP DURATION SIZE TIMESTAMP LEVEL URL PATH

RAND = a high-entropy id (uuid / hash / token); NUM/SIZE/DURATION are numeric values; PUNCT is reserved for separators (so content-token labels draw from the other ten). categories.json carries the per-category evaluation weights and a note on the minimal-set schema.

How it was built

A template is literal text plus typed slots ({TS}, {LEVEL}, {IP}, {RAND}, {NUM}, {DURATION}, {SIZE}, {URL}, {PATH}, {WORD}). The generator fills each slot from a per-type value generator, records the value's character span, then tokenizes the line and gives each content token the tag of the slot covering its start char — so a multi-token value like Jun 14 10:00:00 is entirely TIMESTAMP. Seeded and reproducible (python loop/synth_gold.py). Validation draws from a disjoint template set, so it measures generalization to unseen log structures.

License

Apache-2.0. The data is generated by OpsBR's own template engine and contains no third-party log content — see LICENSE / NOTICE.

Usage

from datasets import load_dataset
ds = load_dataset("opsbr/eye-grep")     # -> {"train": 2943, "validation": 146}
print(ds["validation"][0]["line"], ds["validation"][0]["gold"])