LaurelWings/rcga-lingcoder-training-data

Ling-Coder training snapshots for RCGA / LoopSFT

Derived from inclusionAI/Ling-Coder-SFT, original authors Codefuse and Ling Team, Apache-2.0. The local raw release contained 5,119,470 rows in 26 shards; its recorded upstream revision is 7c631c8eb4f9b72f724a8cc88eb5c19f1cdecfa8. The upstream dataset card and citation are retained in provenance/.

Two different 400k selections are preserved as separate configs. They must not be silently interchanged.

Config	Files / split	Role
dense17_moe30_400k	400,000-row Parquet, train	Exact file used by the current 1.7B and 30B-A3B runs
qwen4b_seed42	400,000 train + 4,096 validation, JSONL	Historical 4B train/dev split, with source identifiers and audit

Current 1.7B / 30B selection

data/dense17_moe30/sft_ling_coder_400k.parquet is copied directly from the file used for training; schema is one messages column with role/content pairs. It is not pre-tokenized. This saved file has no separate validation split; all 400,000 rows are the fixed training pool. Runtime training shuffle is separate from dataset selection.

The archived recipes/prepare_sft_dataset.py normalizes roles, optionally filters Python language tags, rejects malformed/unclosed-think/overlength examples, uses shard striding, and takes a seeded sample. Its original complete preparation command for this exact final file was not retained; defaults alone are not asserted to reproduce identical membership. The already prepared file is therefore the authoritative reproducible split. No filtering or text changes were introduced during backup.

Historical 4B split

The original preparation_audit.json and preparation script are preserved. From the fixed raw dataset, seed42 assigns deterministic BLAKE2b priorities to mid, selects a candidate pool, checks single user/assistant turns, normalizes question text (NFKC/newlines/outer whitespace) for deduplication, and drops duplicate question groups or mids. Accepted conversations must fit complete ChatML within 4096 tokens; answers are not truncated. The first 4096 accepted examples form dev, and the next 400000 form train. The grouping normalization does not rewrite the saved training message text. dev.jsonl is exposed as the HF validation split.

The existing exact train.jsonl is reused from the owner's earlier backup, rather than reselecting examples. Its provenance is recorded in manifest.json. The earlier repo had manual gating; this requested backup is public and ungated.

from datasets import load_dataset
train = load_dataset('LaurelWings/rcga-lingcoder-training-data', 'dense17_moe30_400k', split='train')
old = load_dataset('LaurelWings/rcga-lingcoder-training-data', 'qwen4b_seed42')

Download original files without schema conversion:

from huggingface_hub import snapshot_download
snapshot_download('LaurelWings/rcga-lingcoder-training-data', repo_type='dataset', local_dir='training-backup')

For existing 1.7B/30B launchers, restore the Parquet to /workspace/zy85/datasets/sft_ling_coder_400k.parquet. The raw 5.1M-row upstream corpus was not duplicated; the exact subsets actually trained on are preserved here.