LiamDuero's picture
Add dataset card (README.md)
74a093b verified
|
Raw History Blame Contribute Delete
9.11 kB
---
pretty_name: Telco-Retrieve Chunks
language:
- en
# license: to be decided by GSMA. "other" is a placeholder.
license: other
tags:
- telecom
- 3gpp
- rag
- retrieval
- embeddings
- chunks
---
# Telco-Retrieve Chunks
Pre-built retrieval indexes over chunks of the 3GPP Release 19 specifications: for each configuration a ChromaDB directory (vectors, cosine) and an SQLite FTS5 (BM25) database. They were built with the open-telco-rag pipeline (`Znbne/telco-retrieve`, version 0.1.0, commit `fdda58395243121d52af1f74724c826cd94e66c7`) from the text of `GSMA/3GPP` at revision `a056f6018a7e8e67052aa68a702e272d0ae95d75`. The fixed grid in `telcoretrievechallenge/` is the set of indexes used by the TelcoRetrieve Challenge. You can skip the ingestion step and query an index directly, for example with `open-telco-rag challenge export`.
License, authorship and publication terms: to be decided by GSMA. See also the provenance note recorded in `GRID_MANIFEST.json` (quoted at the end of this card).
## What is in this repo
| Path | What it is | In the challenge? | Files | Size |
|---|---|---|---|---|
| `telcoretrievechallenge/fixed-grid/<strategy>/<embedder>/` | Base grid: 3 chunking strategies x 6 embedders = 18 cells, no enrichment | yes | 216 | 170.40 GB |
| `telcoretrievechallenge/fixed-grid/enrich/<enrichment>/<strategy>/<embedder>/` | Enrichment grid: 2 enrichments x 3 strategies x 6 embedders = 36 cells | yes | 432 | 406.08 GB |
| `telcoretrievechallenge/fixed-grid/large/<model>/...` | 8B-embedder cells: 2 models x (3 + 6) = 18 cells | yes | 290 | 445.93 GB |
| `telcoretrievechallenge/fixed-grid/GRID_MANIFEST.json`, `README.md`, `archive_driver1/` | Manifest, grid description, early driver logs | yes | 5 | under 1 MB |
| `telcoretrievechallenge/README.md` | Short guide to the fixed grid | yes | 1 | small |
| `archive/` | Earlier and experimental material, moved here unchanged | **no** | 1668 | 1,047.74 GB |
| `README.md`, `.gitattributes` | This card; repository configuration | n/a | 2 | small |
`telcoretrievechallenge/` is the fixed grid used by the challenge. `archive/` is **not** part of the challenge and must not be used for it.
## The fixed grid
Every base cell was built with the same settings (from `GRID_MANIFEST.json`):
- Corpus: 3GPP Release 19 only (`GSMA/3GPP`, revision `a056f6018a7e8e67052aa68a702e272d0ae95d75`).
- Distance metric: cosine. Enrichment: none. Chunk bound: 512 tokens (`bound_outliers`: split).
- Chunking: `hierarchical_markdown` `--chunk-size 2048`; `text_baseline` `--chunk-size 1024 --overlap 200`; `sliding_window_tokens` `--chunk-size 512 --overlap 50` (sizes as given to the pipeline).
- Pipeline: `Znbne/telco-retrieve` commit `fdda58395243121d52af1f74724c826cd94e66c7` (`git_dirty`: false). The 8B-embedder cells under `large/` record a different commit (`c5f7886c549169c6891e0e664c934e6f342f3d11`, `git_dirty`: true) and an uncommitted diff stored next to them.
- Collection names: `<embedder>_<strategy>_b512_v256_full` for the base grid; `<embedder>_<strategy>_<enrichment>_b512_v256_full` for the enrichment grid (`<embedder>_<strategy>_<enrichment>_b512_v256_full`).
Base cells (18):
| strategy | embedder | chunks | dim | size | collection |
|---|---|---|---|---|---|
| hierarchical_markdown | minilm | 1,086,324 | 384 | 7.39 GB | `minilm_hierarchical_markdown_b512_v256_full` |
| hierarchical_markdown | bge | 1,086,324 | 1024 | 10.17 GB | `bge_hierarchical_markdown_b512_v256_full` |
| hierarchical_markdown | bge-m3 | 1,086,324 | 1024 | 10.18 GB | `bge-m3_hierarchical_markdown_b512_v256_full` |
| hierarchical_markdown | qwen3-embedding-0.6b | 1,086,324 | 1024 | 10.26 GB | `qwen3-embedding-0.6b_hierarchical_markdown_b512_v256_full` |
| hierarchical_markdown | otel-0.6b | 1,086,324 | 1024 | 10.20 GB | `otel-0.6b_hierarchical_markdown_b512_v256_full` |
| hierarchical_markdown | embeddinggemma-300m-fp32 | 1,086,324 | 768 | 9.17 GB | `embeddinggemma-300m-fp32_hierarchical_markdown_b512_v256_full` |
| text_baseline | minilm | 1,072,821 | 384 | 7.15 GB | `minilm_text_baseline_b512_v256_full` |
| text_baseline | bge | 1,072,821 | 1024 | 9.86 GB | `bge_text_baseline_b512_v256_full` |
| text_baseline | bge-m3 | 1,072,821 | 1024 | 9.88 GB | `bge-m3_text_baseline_b512_v256_full` |
| text_baseline | qwen3-embedding-0.6b | 1,072,821 | 1024 | 9.96 GB | `qwen3-embedding-0.6b_text_baseline_b512_v256_full` |
| text_baseline | otel-0.6b | 1,072,821 | 1024 | 9.90 GB | `otel-0.6b_text_baseline_b512_v256_full` |
| text_baseline | embeddinggemma-300m-fp32 | 1,072,821 | 768 | 8.89 GB | `embeddinggemma-300m-fp32_text_baseline_b512_v256_full` |
| sliding_window_tokens | minilm | 1,094,155 | 384 | 7.40 GB | `minilm_sliding_window_tokens_b512_v256_full` |
| sliding_window_tokens | bge | 1,094,155 | 1024 | 10.16 GB | `bge_sliding_window_tokens_b512_v256_full` |
| sliding_window_tokens | bge-m3 | 1,094,155 | 1024 | 10.18 GB | `bge-m3_sliding_window_tokens_b512_v256_full` |
| sliding_window_tokens | qwen3-embedding-0.6b | 1,094,155 | 1024 | 10.26 GB | `qwen3-embedding-0.6b_sliding_window_tokens_b512_v256_full` |
| sliding_window_tokens | otel-0.6b | 1,094,155 | 1024 | 10.20 GB | `otel-0.6b_sliding_window_tokens_b512_v256_full` |
| sliding_window_tokens | embeddinggemma-300m-fp32 | 1,094,155 | 768 | 9.18 GB | `embeddinggemma-300m-fp32_sliding_window_tokens_b512_v256_full` |
Enrichment grid: `metadata_tagging` and `acronym_expansion`, applied after the bound step, so enriched chunks can exceed the token bounds (see `enrichment_note` in `GRID_MANIFEST.json`). Large cells: `Qwen/Qwen3-Embedding-8B` (`qwen3-embedding-8b-local`, 4096 dimensions) and `nvidia/Nemotron-3-Embed-8B-BF16` (`nemotron-3-embed-8b-local`, 4096 dimensions), embedded locally.
**Default embedder of `challenge export`.** The default profile of `open-telco-rag challenge export` uses `--model otel-0.6b` (registry: `farbodtavakkoli/OTel-Embedding-0.6B`, 1024 dimensions). This matches the manifest: the `otel-0.6b` cells have `vector_dim` 1024. The default cell to start with is `telcoretrievechallenge/fixed-grid/text_baseline/otel-0.6b`.
## How to use it
Download one cell (not the whole repo):
```bash
hf download GSMA/telco-retrieve-chunks --repo-type dataset \
--include "telcoretrievechallenge/fixed-grid/text_baseline/otel-0.6b/*" --local-dir ./runs
```
Then export a submission:
```bash
open-telco-rag challenge export --questions Test.csv --out submission.csv \
--db-path ./runs/telcoretrievechallenge/fixed-grid/text_baseline/otel-0.6b/db \
--collection otel-0.6b_text_baseline_b512_v256_full \
--generator-model <served-model-name> --llm-base-url http://localhost:8000/v1
```
`--collection` is required and has no default in the CLI. The name above is the one used in the CLI's own help example and it is in `GRID_MANIFEST.json`. These commands were checked against `hf download --help` and `open-telco-rag challenge export --help` (from the source of `challenge-telco-retrieve` at commit 51c7211). They were not run end to end.
## Archive
`archive/` holds earlier and experimental material: the 120 root folders this repo had before (120 folders `3gpp-r19_<embedder>_<chunking>_<enrichment>_<params>/`), the previous README and a copy of the previous `.gitattributes`. It is kept for reference. It is a different set from the fixed grid, it is not part of the challenge and it is not maintained. See `archive/ARCHIVE_NOTE.md`; the path rule is `<old path>` to `archive/<old path>`.
## Integrity and provenance
- `telcoretrievechallenge/fixed-grid/GRID_MANIFEST.json` records, for each of the 54 base and enrichment cells, every file with its size and sha256. The 18 `large/` cells are described in the manifest (collection, chunks, bytes) but their files are not listed individually.
- Check a downloaded cell against the manifest (run next to `./runs`; it skips files you did not download):
```python
import hashlib, json, pathlib
r = pathlib.Path("runs/telcoretrievechallenge")
m = json.load(open(r / "fixed-grid/GRID_MANIFEST.json")) # also download this file
cells = m["cells"] + m["enrichment_grid"]["cells"]
for c in cells:
for rel, v in c["files"].items():
p = r / c["path_in_repo"] / rel
if p.exists():
print(c["path_in_repo"], rel, hashlib.file_digest(open(p, "rb"), "sha256").hexdigest() == v["sha256"])
```
- Build commit: `fdda58395243121d52af1f74724c826cd94e66c7` in `Znbne/telco-retrieve`. Code and documentation: https://github.com/Znbne/telco-retrieve
- Provenance note from `GRID_MANIFEST.json` (`licence_note`): "Provenance and terms: the source documents are 3GPP specifications and remain the property of 3GPP and its Organizational Partners (see https://www.3gpp.org). The text in these collections is chunked from the GSMA/3GPP dataset (https://huggingface.co/datasets/GSMA/3GPP, revision a056f6018a7e8e67052aa68a702e272d0ae95d75), which describes itself as mirroring the documents for research, search and retrieval. The collections contain verbatim chunk text. Use them subject to the original 3GPP terms; this repository grants no additional rights."