dirac-run/ec-training-data

EasyCommand training data

401,975 unique English request/Bash command pairs for training GNU/Linux shell command generators. The release is 167,983,166 bytes (160.2 MiB), with 154,807 distinct command strings. Different descriptions of the same command are intentionally retained; there are no exact duplicate request/response pairs.

Format and loading

The single train.jsonl file has one record per line:

{"request":"list files in this directory","response":{"kind":"COMMAND","value":"ls"}}

All responses use kind: COMMAND. There are no clarification or inability categories. response.value is the Bash command, not an instruction to execute the record while loading or preprocessing it.

from datasets import load_dataset

data = load_dataset("dirac-run/ec-training-data", split="train")
print(data[0]["request"])
print(data[0]["response"]["value"])

To load a downloaded file locally:

data = load_dataset("json", data_files="train.jsonl", split="train")

Only a training split is supplied. There is no official validation/test split. Create holdouts by command intent/family before making paraphrases; random row splits can place descriptions of the same command on both sides.

Content and construction

Coverage includes GNU/Linux file operations, discovery/find, Git, text processing, archives, processes/system inspection, networking and HTTP, quoting, operands and composed pipelines. Records target Bash and GNU utilities, not every shell or OS.

The project produced synthetic examples through AI-assisted authoring and programmatic generation, with semantic review and selected execution-backed checks. Later work added concise descriptions, repaired labels and scope/argument distinctions, simplified unnecessarily complex commands and added contrast pairs and replay examples. Not every row was individually executed or freshly reviewed again for this export; the dataset is not a guarantee of command correctness.

The published union contains 400,185 distinct pairs from the weighted full-run corpus plus 1,790 new pairs from incremental repair corpora. Historical full-run materialization had 474,635 presentations because some pairs were repeated or reweighted. Exact pair duplicates and metadata were removed for publication. The export preserves different descriptions and includes repair data from experiments beyond the two selected model checkpoints. One flat pass over this file does not reproduce the original training exposure or repair/replay sampling.

Model training format

Use a system message, the request as a user message, and compact serialized response JSON as the assistant target. Supervise the assistant answer/EOS tokens, masking system/user/padding tokens. Disable thinking when using Qwen3.

The selected models use this serving/continuation system message:

You are a GNU/Linux shell command generator. Produce the simplest Bash command that fulfills the entire request. Return only valid JSON: {"kind":"COMMAND","value":"<command>"}.

Earlier parent training used a longer prompt, included in the model repositories. The ec training guide describes formatting, retention pilots and quantization checks. Published models: 0.6B trainable weights/adapters, 1.5B trainable weights/adapters, 0.6B GGUFs and 1.5B GGUF.

Privacy and evaluation boundaries

The export contains only request and response. Source metadata, local research paths, reviewer identities and raw experiment receipts are excluded. The project's personal home prefix was replaced with ~/ in the export; original audit sources were retained privately. Synthetic examples may contain fixture usernames, paths, addresses, IPs and credential-shaped schema/regex text. These are not collected contact records. Pattern/context screening found no confirmed live credentials or genuine personal records, and re-screening this exact export found no new unreviewed values. This is a screening result, not proof that every possible identifier or encoded secret is absent.

ALFA failures and internal benchmarks informed repairs and model selection; the reported model scores are development evidence, not an untouched independent test. The export has zero exact or case/whitespace-normalized matches to the 300 canonical ALFA-updated request strings. That narrow check does not establish absence of task/template overlap or benchmark-informed development. Evaluate generalization on fresh intents, operands, wording and environments.

License and integrity

The project's dataset and this documentation are released under Apache-2.0, with project attribution in NOTICE. The model weights and application have their own repository licenses.

train.jsonl SHA256: 4cc01d4511777cc704f0b5b30dee59e721e41d3c0b3dc3ab47cdb0ecb62051ce. Verify this folder with sha256sum --check SHA256SUMS. manifest.json records schema, counts and file identity.