Datasetpublic

Jun1801/telcollm-oran-training

TelcoLLM ORANBench training data Training inputs for a Qwen3-8B run against ORANBench. The code lives in git; this repo carries only the data, which is too large for it. Generated 2026-09-03. Contents path size what it is train_ready_v1/ 162 MB cleaned Product A and B, with true token counts -- the inputs to every mix replay/ 19 MB in-domain and general replay, already through the split firewall gate splits_v2/ 0 MB the document-family split every… See the full description on the dataset page: https://huggingface.co/datasets/Jun1801/telcollm-oran-training.

sourceHugging Faceotherupdated 2d agoView on Hugging Face

0likes73downloads

Dataset Card

TelcoLLM ORANBench training data

Training inputs for a Qwen3-8B run against ORANBench. The code lives in git; this repo carries only the data, which is too large for it.

Generated 2026-09-03.

path	size	what it is
`train_ready_v1/`	162 MB	cleaned Product A and B, with true token counts -- the inputs to every mix
`replay/`	19 MB	in-domain and general replay, already through the split firewall gate
`splits_v2/`	0 MB	the document-family split every stage is checked against
`sft_set_v1/`	86 MB	rendered SFT / RL / dev sets, document-disjoint
`cpt_mix_b_v1/`	210 MB	the packed CPT mix for arm B -- what training reads directly

Licensing — read before making this public

Everything here is derived from O-RAN specifications. Their terms are distinct from ORANBench's MIT license, and a public download does not imply a right to redistribute. AGENTS.md requires the applicable source-document terms to be verified and documented before publishing derived data. This repo is intended to be private: transfer between machines, not redistribution.

The replay slices carry their own, verified licences: IETF Trust §4.c, 3GPP TDoc (free redistribution), CC0, CC-BY, CC-BY-SA, US/EU public domain (GSMA/Telco-Common-Corpus), and ODC-By (HuggingFaceFW/fineweb-edu).

The evaluation firewall

ORANBench is not in this repo, deliberately. It was used only to exclude contamination during generation — 1,835 of 95,691 candidate MCQs (1.9%) — and never to train, tune or select. Do not add it here.

Product A carries no benchmark leakage: 0 verbatim benchmark questions across 54,409 passages, max TF-IDF cosine 0.6986 against the 1,500 benchmark rows, nothing at 0.80.

Provenance

Every directory ships the manifest.json it was produced with: input hashes, code hash, seeds, tokenizer revision, and the counts each stage kept and dropped. Those are the record; this card is a summary.

Usage

bash

hf download Jun1801/telcollm-oran-training --repo-type dataset --local-dir data/hf

Then point configs/training/*.yaml at the downloaded paths, or move the directories into place:

data/hf/train_ready_v1  -> data/derived/train_ready_v1
data/hf/replay          -> data/derived/replay
data/hf/splits_v2       -> data/derived/splits_v2
data/hf/sft_set_v1      -> artifacts/sft_set/v1
data/hf/cpt_mix_b_v1    -> artifacts/cpt_mix/b_v1

The two control mixes (no_replay, raw_only) are not uploaded — they rebuild deterministically from train_ready_v1 with scripts/build_cpt_mix.py, and a rebuild was verified byte-identical.