CCRss/agents-as-jds-llm-data

Agents as JDS — LLM modality data

Evaluation and training data for an agentic domain-adaptation study. Code, docs and the full explanation live in the companion repo: github.com/vladimiralbrekhtccr/agents-as-jds-llm — read HANDOFF.md there first.

Base model throughout: Qwen3-1.7B (thinking).

Layout

spider/          domain A — text-to-SQL (working)
  target_dev.jsonl        275 rows / 20 dbs
  target_test.jsonl       550 rows / 40 dbs
  pred_dev_base.jsonl     base-model predictions
  pred_test_base.jsonl

kernelbench/     domain B — CUDA kernels (kept, NOT proven)
  v2/target_dev.jsonl      98 rows / 15 units   <- recommended split
  v2/target_test.jsonl    172 rows / 31 units
  v1/                     no-judgement baseline variant
  acceptance/             pool, training set, and the failed fine-tune run

general_mmlu/    shared by both domains
  general_dev_sci.jsonl   500 / 4 subjects
  general_dev_hum.jsonl   500 / 4 subjects
  general_test_sci.jsonl  500 / 4 subjects
  general_test_hum.jsonl  500 / 4 subjects
  general_pool.jsonl      11,346 / 41 subjects   (replay)
  subject_acc.json        per-subject base competence, all 57 subjects
  base_ll.json            measured base error

Base error (Qwen3-1.7B, single pass)

set	error_pct
spider target dev	28.0
spider target test	23.8
general dev sci / hum	33.8 / 26.4
general test sci / hum	35.4 / 26.8

error_pct = 100 - accuracy, lower is better. Chance on MMLU is 75.0.

Two things to know before using this

1. KernelBench is not a validated domain. Its correctness metric is gameable — returning the reference PyTorch scores ~26% while writing zero CUDA kernels. See AUDIT.md in the code repo. Spider's metric passes cheater controls (0.0–0.7%).
2. Never train on the evaluation sets, and never on anything sharing their split unit — db_id for Spider, subject for MMLU. Per-dataset forbidden lists are in the code repo.

Row format

{"id", "input_path", "target", "meta": {"domain", "split_unit", ...}}

Deviations, both deliberate: MMLU rows carry the prompt inline in input with input_path: null (a file-per-row would mean ~16k tiny files); KernelBench rows have target: null because the task has no reference answer string — correctness is decided by execution, not string match.