samples/README.md

OTel 2.5 sample training rows (10 per category, for review before scaling)

Format follows OTel_data_label.md: one JSON object per line with source_repo, source_id (chunk id), source_spec_id: null, source_version: null, and exactly two messages (user = document chunk + question, assistant = answer grounded in that chunk).

file	category	source collection	data origin
lte-signalling.jsonl	LTE S1AP/NAS control-plane events at the EPC	powder-lte-v1	measured, POWDER real radios (srsRAN), benign vs fake-base-station runs
rf-signal-quality.jsonl	LTE measurement reports (RSRP/RSRQ/SINR/PHR/TA, neighbours)	powder-lte-v1	measured, POWDER real radios
5g-core-signalling.jsonl	NGAP, PFCP and SBI request flows in Open5GS	zmq-5g-core-sim-v1, powder-5g-core-sim-v1	measured core signalling; radio simulated (ZMQ) or UERANSIM
ran-telemetry.jsonl	srsUE per-second radio metrics + ocudu gNB scheduler metrics	zmq-5g-core-sim-v1	simulated radio link (ZMQ)
network-performance.jsonl	iperf3/ping/HTTP results and procedure timings per run	zmq-5g-core-sim-v1	measured application results over a simulated link
faults-ground-truth.jsonl	injected NF restarts/stops/config caps (ground-truth timeline) with NF warning/error logs	zmq-5g-core-sim-v1, powder-5g-core-sim-v1	measured logs, ground truth from orchestrator
aaa-radius.jsonl	FreeRADIUS Access/Accounting exchanges	powder-5g-core-sim-v1	measured

How they were made (pipeline/trainrows/ in the collection repo):

1. render.py turns processed, pseudonymised records of one run into a plain-text document with a header (run id, scenario label, provenance statement, ground-truth timeline for fault runs) and a record table.
2. Documents are chunked on line boundaries (header repeated), sized so the largest row stays under ~3.5k tokens in the generator tokenizer (tables tokenize at ~2 chars per token), under the 4,096-token row limit.
3. generate.py asks google/gemma-4-31b-it (via OpenRouter) for one question an operations engineer would ask about that chunk and an answer using only the chunk; a second judge call checks every claim is supported and that provenance is not over-claimed (simulated radio never presented as over-the-air). Rejected rows are dropped.
4. Rows are scanned for raw identifiers (15-digit numbers, test IMSIs); subscriber ids are keyed pseudonyms.

Each *.log.json lists, per generated row, the run, scenario, judge verdict and token usage. Rows are sampled round-robin over scenarios so each category covers several conditions. Known limits of this first batch: questions are single-chunk and mostly descriptive (counts, sequences, before/after comparisons); the judge model is the same model as the generator; chunks are tables, not prose.