Create README.md

README.md

@@ -0,0 +1,198 @@
+---
+license: apache-2.0
+language:
+  - en
+pretty_name: Telco-Retrieve QnA (3GPP)
+task_categories:
+  - question-answering
+  - multiple-choice
+  - text-generation
+tags:
+  - telecom
+  - telecommunications
+  - 3gpp
+  - rag
+  - retrieval-augmented-generation
+  - evaluation
+  - benchmark
+  - synthetic
+  - ground-truth
+configs:
+  - config_name: question_pool
+    data_files: question_pool_3gpp.json
+  - config_name: set_840
+    data_files: 3gpp_question_set_840.json
+  - config_name: set_105
+    data_files: 3gpp_question_set_105.json
+  - config_name: set_84
+    data_files: 3gpp_question_set_84.json
+# NOTE: records live under the "questions" key of each JSON object. If the dataset
+# viewer should expand them, either add a loading script or flatten to JSONL.
+---
+
+# Telco-Retrieve QnA (3GPP)
+
+Synthetic, expert-validated question–answer pairs over **3GPP** standards, built to evaluate
+retrieval-augmented generation (RAG) and closed-book LLM performance on telecom domain knowledge.
+Part of the **[GSMA Open Telco AI](https://www.open-telco.ai/)** initiative and the companion
+evaluation set for [`GSMA/telco-retrieve-chunks`](https://huggingface.co/datasets/GSMA/telco-retrieve-chunks)
+and the [`open-telco-rag`](https://github.com/Znbne/telco-retrieve) pipeline.
+
+Every question is grounded in a specific chunk of a real 3GPP document, carries a canonical spec
+citation, and has passed a multi-stage LLM jury followed by human review.
+
+## Dataset Summary
+
+A validated pool of **1,227** 3GPP questions plus three pre-drawn, stratified evaluation splits of
+different sizes. Questions come in three formats — multiple-choice (`mc`, 2–5 options), true/false
+(`tf`), and open-ended (`open`) — and are filtered so that every retained question is answerable
+*from its source chunk*, discriminative against similar chunks, and non-trivial without retrieval.
+
+Intended as a **held-out benchmark, not training data**.
+
+- **Corpus:** 3GPP — Release 8 through Release 19, all series (TS 23/28/32/33/38 …)
+- **Source chunks:** [`GSMA/telco-retrieve-chunks`](https://huggingface.co/datasets/GSMA/telco-retrieve-chunks) (`3gpp` export)
+- **Generation:** [`telco-qna-generation`](https://github.com/Znbne/telco-qna-generation) declarative Flows
+- **All records:** `validated: true` (expert-reviewed)
+
+## Files
+
+| File | Config | Questions | MC / TF / OE | Size | What it is |
+|---|---|---:|---|---:|---|
+| `question_pool_3gpp.json` | `question_pool` | 1,227 | 475 / 514 / 238 | 3.6 MB | The full validated question pool. Draw your own samples from this. |
+| `3gpp_question_set_840.json` | `set_840` | 840 | 400 / 400 / 40 | 2.4 MB | Large stratified evaluation sample (difficulty-first + 3GPP-series-stratified). |
+| `3gpp_question_set_105.json` | `set_105` | 105 | 55 / 40 / 10 | 317 KB | Small stratified sample; highest-difficulty preferred, series-stratified. |
+| `3gpp_question_set_84.json` | `set_84` | 84 | 40 / 40 / 4 | 254 KB | Independent 84-question draw (40/40/4), same stratification. |
+
+The three `*_question_set_*` files are **subsets drawn from `question_pool_3gpp.json`** — they
+overlap with the pool and, partially, with each other. Pick one split as your benchmark; don't sum
+row counts across files.
+
+> `question_pool_3gpp.json`'s top-level `_generated` header still reads "Requires expert validation
+> before use as ground truth" — that string is a stale default. Every record in it has
+> `validated: true`; trust the per-record field, not the file header.
+
+## Dataset Structure
+
+Each file is a single JSON object:
+
+```json
+{
+  "_corpus": "3gpp",
+  "_description": "…how this split was drawn…",
+  "_generated": "2026-08-27T20:08:55Z",
+  "questions": [ { …record… }, … ]
+}
+```
+
+### Record fields
+
+| Field | Type | Notes |
+|---|---|---|
+| `question_id` | string | 12-hex stable id (join key for review sheets) |
+| `question` | string | Question text |
+| `reference_answer` | string | Gold answer. `mc`: the correct option; `tf`: `"true"`/`"false"`; `open`: short free-text answer |
+| `answer_type` | string | `mc` \| `tf` \| `open` |
+| `options` | object | **MC only** — map of option label → option text |
+| `explanation` | string | **TF only** — why the statement is true/false |
+| `difficulty` | string | Model-assigned at generation (`unspecified` if not set) |
+| `keywords` | list[string] | Salient terms, for keyword-overlap retrieval metrics |
+| `corpus` | string | Always `3gpp` in this dataset |
+| `canonical_source` | string | Human-readable citation, e.g. `TS 28.312 (Rel-18)` |
+| `release` | string | 3GPP release the source chunk was ingested from (`""` where N/A) |
+| `spec_version` | string | Full spec version string (present in some sets) |
+| `chunk_id` | string | Exact source chunk in `GSMA/telco-retrieve-chunks` |
+| `source_document` | string | Source file the chunk came from |
+| `source_text` | string | **Verbatim excerpt** of the 3GPP chunk the question was written from |
+| `generated` | bool | Always `true` (all items model-generated) |
+| `validated` | bool | `true` — expert-reviewed and merged into the canonical set |
+| `validation_passed` | bool | Answerability / self-consistency check |
+| `validation_note` | string | Note from the answerability check |
+| `domain_relevance_passed` | bool | LLM jury: is this a telecom-domain question |
+| `faithfulness_passed` | bool | LLM jury: answer entailed by `source_text` |
+| `semantic_correctness_passed` | bool | LLM jury: answer is semantically correct |
+| `relevance_passed` | bool | LLM jury: question relevant to the chunk |
+| `competitor_chunks` | int | # distractor chunks in the discriminativeness check |
+| `competitor_correct_ratio` | float | Fraction of distractor chunks that also "answer" the question (lower = more discriminative) |
+| `competitor_threshold_used` | float | Pass threshold applied to `competitor_correct_ratio` |
+| `jury_no_context_score` | float | Empirical-difficulty jury score with **no** retrieved context |
+| `jury_with_context_score` | float | Empirical-difficulty jury score **with** context |
+| `no_context_threshold_used` | float | Threshold separating too-easy from rag-suitable |
+| `empirical_difficulty` | string | e.g. `rag_suitable`, `too_easy`, `unanswerable` |
+| `review_status` | string | Human review outcome |
+| `reviewer` | string | Reviewer id |
+| `review_notes` | string | Human reviewer notes |
+
+## Supported Tasks
+
+- **RAG evaluation** — retrieve over `GSMA/telco-retrieve-chunks` (`3gpp`), answer, score against
+  `reference_answer` / `options`, and check retrieval with `canonical_source` (canonical hit rate).
+- **Closed-book QA** — no-retrieval baseline: parametric 3GPP knowledge and abstention behaviour.
+- **Multiple-choice accuracy** and **open-ended answer correctness** (LLM-graded vs `reference_answer`).
+
+The `open-telco-rag evaluate` harness consumes a split directly as
+`configs/eval_questions/3gpp.json`.
+
+## Source Data & Curation
+
+### Source
+
+3GPP specifications (Rel-8 → Rel-19, all series), ingested and chunked in
+`GSMA/telco-retrieve-chunks`. Underlying 3GPP text was sourced via the TSpec-LLM corpus and the
+3GPP spec archive (`3gpp.org/ftp/Specs/archive`).
+
+### Generation pipeline
+
+Produced by [`telco-qna-generation`](https://github.com/Znbne/telco-qna-generation) —
+declarative YAML **Flows** chaining composable blocks:
+
+1. **Sample chunks** — stratified over the corpus by 3GPP TS series.
+2. **Generate questions** — per-format prompts with 3GPP-specific guidance, written from the source
+   text (not a summary).
+3. **Validate & dedup** — answerability check; prefix-dedup against prior drafts.
+4. **LLM jury** — grounding, answerability, faithfulness, domain-relevance, semantic-correctness and
+   relevance blocks; each records a `*_passed` flag.
+5. **Competitor filter** — re-asks the question against distractor chunks; keeps only questions a
+   reader can't answer from the wrong chunk.
+6. **Empirical-difficulty jury** — scores the question with and without context; drops `too_easy`
+   and `unanswerable`, keeps `rag_suitable`.
+7. **Draft → human review → merge** — output is a *draft* (`generated: true, validated: false`)
+   until a domain expert reviews it and it is explicitly merged into the canonical set
+   (`validated: true`).
+
+### Evaluation splits
+
+`question_pool_3gpp.json` is the merged, validated pool. The `*_question_set_*` files are drawn
+from it **difficulty-first** (highest-difficulty questions preferred) and **stratified by 3GPP
+series** (from the `chunk_id` prefix) to preserve topical spread. `set_105` and `set_84` are
+independent draws; `set_84` is the 40/40/4 subset used for the thesis's per-question analysis.
+
+### Curation rationale
+
+The aim is a benchmark that isolates *retrieval-conditioned* 3GPP knowledge: questions must be
+answerable from a single cited chunk, hard to answer without it, and discriminative against similar
+chunks — so RAG lift over a closed-book baseline is measurable rather than confounded by questions
+a model already knows or cannot answer at all.
+
+## Considerations & Limitations
+
+- **Synthetic questions, LLM-assisted validation.** Every question is model-generated and most
+  filters are LLM juries. Human review is a final gate but is single/limited-reviewer, so residual
+  errors in `reference_answer` or `canonical_source` are possible. `validated: true` means
+  "expert-checked", not "formally adjudicated".
+- **Single corpus.** 3GPP only. Not representative of O-RAN, ETSI, ITU-T, CAMARA, GSMA or TM Forum,
+  which the `telco-qna-generation` toolkit can target but this dataset does not cover.
+- **Point-in-time.** Tied to specific spec releases (`release` / `canonical_source`); 3GPP moves on.
+- **Not training data.** Held-out evaluation only. Do not fine-tune on it or leak it into training.
+- **Grader dependence.** OE scoring uses an LLM grader against `reference_answer`; absolute OE
+  numbers depend on that grader.
+- **English only.**
+
+## Licensing
+
+- **QnA annotations** (questions, answers, keywords, judge/review metadata): **Apache-2.0**,
+  consistent with the `telco-qna-generation` toolkit. *(Set here as `apache-2.0`; change if GSMA
+  requires a different licence for this repo.)*
+- **`source_text`** contains short verbatim excerpts of 3GPP specifications. Those excerpts remain
+  subject to 3GPP's copyright and redistribution terms and those of the upstream source
+  (the TSpec-LLM dataset / the 3GPP spec archive). Downstream users are responsible for compliance.