Release v0.9.0: Modern AI & Systems (2020-2026) multi-task dataset

README.md

@@ -1,464 +1,174 @@
----
-language:
-  - en
-license: other
-license_name: mixed-per-record-community-content
-license_link: https://creativecommons.org/licenses/
-pretty_name: Community QA Lab v0.7.0 Multisource RAG Corpus
-size_categories:
-  - 10K<n<100K
-tags:
-  - datasets
-  - text
-  - community-content
-  - retrieval
-  - rag
-  - answer-reranking
-configs:
-  - config_name: default
-    default: true
-    data_files:
-      - split: train
-        path: train.parquet
-      - split: validation
-        path: validation.parquet
-      - split: test
-        path: test.parquet
-  - config_name: legacy_default_v0.4.0
-    data_files:
-      - split: train
-        path: releases/v0.4.0/train.parquet
-      - split: validation
-        path: releases/v0.4.0/validation.parquet
-      - split: test
-        path: releases/v0.4.0/test.parquet
-  - config_name: technical_qa
-    data_files:
-      - split: train
-        path: technical_qa/v0.5.0/train.parquet
-      - split: validation
-        path: technical_qa/v0.5.0/validation.parquet
-      - split: test
-        path: technical_qa/v0.5.0/test.parquet
-  - config_name: corpus_rag_multisource_merged_v0.1
-    data_files:
-      - split: train
-        path: v0.7.0-multisource-merged/views/corpus_rag_multisource_merged_v0.1/train.parquet
-      - split: validation
-        path: v0.7.0-multisource-merged/views/corpus_rag_multisource_merged_v0.1/validation.parquet
-      - split: test
-        path: v0.7.0-multisource-merged/views/corpus_rag_multisource_merged_v0.1/test.parquet
-  - config_name: preference_dpo_multisource_merged_v0.1
-    data_files:
-      - split: train
-        path: v0.7.0-multisource-merged/views/preference_dpo_multisource_merged_v0.1/train.parquet
-      - split: validation
-        path: v0.7.0-multisource-merged/views/preference_dpo_multisource_merged_v0.1/validation.parquet
-      - split: test
-        path: v0.7.0-multisource-merged/views/preference_dpo_multisource_merged_v0.1/test.parquet
-  - config_name: reranking_listwise_multisource_merged_v0.1
-    data_files:
-      - split: train
-        path: v0.7.0-multisource-merged/views/reranking_listwise_multisource_merged_v0.1/train.parquet
-      - split: validation
-        path: v0.7.0-multisource-merged/views/reranking_listwise_multisource_merged_v0.1/validation.parquet
-      - split: test
-        path: v0.7.0-multisource-merged/views/reranking_listwise_multisource_merged_v0.1/test.parquet
-  - config_name: issue_resolution_multisource_merged_v0.1
-    data_files:
-      - split: train
-        path: v0.7.0-multisource-merged/views/issue_resolution_multisource_merged_v0.1/train.parquet
-      - split: validation
-        path: v0.7.0-multisource-merged/views/issue_resolution_multisource_merged_v0.1/validation.parquet
----
-
-# Community QA Lab
-
-Community QA Lab is a reproducible, multisource community-content corpus. Its
-primary use is NLP/RAG over questions, answers, discussions, issues and related
-technical problem-solving content from public platforms. The default dataset
-at this repository root is the merged v0.7.0 RAG corpus.
-
-The release also contains optional derived views for ranking experiments. Those
-views are not the definition of the corpus, and their interaction fields are
-not factual ground-truth labels.
-
-## Public dataset
-
-The published multisource bundle contains:
-
-- `corpus_rag_multisource_merged_v0.1`: 39,215 licensed post documents;
-- `reranking_listwise_multisource_merged_v0.1`: 4,550 question groups;
-- `preference_dpo_multisource_merged_v0.1`: 7,453 derived weak-preference pairs;
-- `issue_resolution_multisource_merged_v0.1`: 1,823 observed Issue/PR records.
-
-The RAG view is the primary content corpus. Each document keeps question
-context, post text and provenance metadata. `accepted`, `has_explicit_answer`
-and lifecycle fields describe platform observations; they do not mean that the
-text has been verified as factually correct. Use `document_text` for retrieval
-input and use provenance fields to identify the source.
-
-```python
-from datasets import load_dataset
-
-rag = load_dataset("aluha501/community-qa-lab")
-```
-
-This one-argument loader returns the merged RAG corpus (`default`): 39,215
-licensed documents across the `train`, `validation` and `test` splits. The
-named ranking and lifecycle views remain available as optional versioned
-artifacts, but they are not required for ordinary NLP/RAG use.
-
-### Historical compatibility configurations
-
-The immutable v0.4.0 files remain available under the explicit
-`legacy_default_v0.4.0` configuration. The v0.5.0 technical QA files remain
-available through the opt-in `technical_qa` config. This keeps the historical
-Parquet artifacts recoverable without making them the default corpus.
-
-- one simple current corpus configuration, `default`, plus opt-in historical/configured views;
-- three Parquet splits: `train`, `validation`, `test`;
-- 10,206 answer rows from 4,039 question groups;
-- AI, Data Science, and Cross Validated (`stats.stackexchange.com`);
-- release date: `2026-08-10`, from an archived API snapshot completed at
-  `2026-08-09T20:59:37.734430Z`;
-- 11 stable columns, including per-answer attribution and update time.
-
-`technical_qa` uses the same 11 columns but is a separate domain benchmark:
-
-- 4,924 answer rows from 1,880 question groups;
-- Software Engineering, Code Review, Quantum Computing, Scientific Computing,
-  and Theoretical Computer Science;
-- strict temporal split: train through 2022, validation in 2023, test from
-  2024 onward;
-- 342 test groups, all refreshed from an archived API snapshot completed at
-  `2026-08-10T02:59:33.978680Z`.
-
-The release retains only threads with at least two candidate answers and
-exactly one accepted answer. It uses a strict question-level temporal split:
-train through 2020-12-31, validation during 2021, and test from 2022 onward.
-The resulting test set has 719 groups, exceeding the release gate of 500.
-These are compatibility/ranking configurations, not replacements for the
-multisource content corpus. Group rows by `question_id` before computing any
-ranking metrics.
-
-```python
-from datasets import load_dataset
-
-dataset = load_dataset("aluha501/community-qa-lab")
-technical = load_dataset("aluha501/community-qa-lab", "technical_qa")
-```
-
-For the legacy flat configurations, a ranker should use `question` and
-`answer` as input. Use `accepted`, `vote_score`, and `disagreement_type` only
-for evaluation or analysis. For v0.7 RAG, use only the text fields as model
-input; IDs, provenance and interaction fields are metadata.
-
-For explicit training and listwise evaluation contracts, build the separate
-local derived views described in [`docs/ranking_views_v0.1.md`](docs/ranking_views_v0.1.md):
-
-```powershell
-python scripts/build_ranking_views.py `
-  --input-dir data/processed/v0.5.0/technical_qa `
-  --output-dir data/derived/ranking_views/v0.1/technical_qa
-```
-
-This creates `ranking_grouped_v0.1` and `ranking_pairs_v0.1` outputs without
-modifying the public release or publishing remote artifacts.
-
-## Expanded Community QA v0.6.0 (local-first)
-
-The experimental expanded config
-[`configs/dataset/technical_qa_v0.6.0_expanded.yaml`](configs/dataset/technical_qa_v0.6.0_expanded.yaml)
-covers 38 technical communities. Its API seed collector is deterministic and
-stratified across train (through 2022), validation (2023), and test (from
-2024). The default pilot is one page per tag for `cs`, `math`, and `devops`;
-the full profile is intentionally a separate command. Responses are cached
-with request metadata and SHA-256 hashes, and a cache rerun must make zero
-network requests. There is no HTML fallback.
-
-Legacy roots are ID sources only. API responses supply all released text,
-attribution, license and metadata; `scraped_cv_legacy` is not automatically
-included. Comments are deferred. Accepted status is an author-interaction
-weak label and votes are community-interaction diagnostics, never factual
-truth.
-
-Build the local views after a completed API snapshot:
-
-```powershell
-python -m src.community_qa_lab.cli data api-seed `
-  --config configs/dataset/technical_qa_v0.6.0_expanded.yaml `
-  --output-dir data/raw/api/v0.6.0-expanded-seeds-pilot `
-  --profile pilot
-
-python scripts/build_expanded_views.py `
-  --release-dir data/processed/v0.6.0-expanded `
-  --config configs/dataset/technical_qa_v0.6.0_expanded.yaml `
-  --cache-dir data/raw/api/v0.6.0-expanded `
-  --output-dir data/derived/v0.6.0-expanded
-```
-
-This produces `preference_dpo`, `reranking_listwise`, and `corpus_rag` as
-local-only Parquet views. See
-[`docs/releases/v0.6.0-expanded-local.md`](docs/releases/v0.6.0-expanded-local.md)
-for the complete audit, enrichment, validation and safe staging workflow.
-
-## Multisource v0.7.0 (experimental; merged derivative published)
-
-The v0.7 derivative is a read-only multisource content corpus. It includes
-Stack Exchange text, GitHub Discussions/Issues/Pull Requests, and a small set
-of GitLab retrieval/resolution records from explicitly configured sources.
-The primary release view is `corpus_rag_multisource_merged_v0.1`, which contains
-licensed documents for retrieval; it is not a factuality benchmark.
-
-The implementation workflow is documented in
-[`docs/releases/v0.7.0-multisource-local.md`](docs/releases/v0.7.0-multisource-local.md)
-and produces the content corpus plus optional views for marked-answer
-preference pairs, listwise Discussion groups and issue-resolution metadata. It
-does not change the v0.4.0/v0.5.0/v0.6.0 artifacts.
-
-The validated merged derivative is now published in the existing
-[`aluha501/community-qa-lab`](https://huggingface.co/datasets/aluha501/community-qa-lab)
-dataset under four opt-in configurations. The publication record and exact
-Hub commit are in
-[`docs/releases/v0.7.0-multisource-merged.md`](docs/releases/v0.7.0-multisource-merged.md).
-
-The DPO/listwise fields are optional weak supervision derived from platform
-interaction (`accepted`/marked answer), not ground-truth correctness labels.
-`issue_resolution` records observed lifecycle/evidence. If you want a normal
-NLP/RAG corpus, load only `corpus_rag_multisource_merged_v0.1` and use
-`document_text`; the other fields are provenance or diagnostics.
-
-After the bounded pilot passed, the larger local sample uses
-`configs/dataset/github_v0.7.0_multisource_large.yaml` with six repositories
-and 25 records per source kind/repository. It remains experimental and is
-kept separate from the small pilot artifacts.
-
-The extended local sample uses
-`configs/dataset/github_v0.7.0_multisource_extended.yaml` with up to 100
-Issues and 100 Discussions per repository. The verified snapshot contains
-904 threads, 3,537 canonical posts and 2,999 RAG documents; it remains a
-local experimental corpus, not a temporally balanced benchmark.
-
-The current scale snapshot uses
-`configs/dataset/github_v0.7.0_multisource_scale.yaml` with 23 explicit
-repositories. It contains 3,018 threads, 12,449 canonical posts, 9,848 RAG
-documents, 253 listwise groups and 1,047 preference pairs. See the
-[`v0.7.0 scale EDA report`](reports/local/v0.7.0-multisource-scale/eda_summary.md)
-and [`local-only release note`](docs/releases/v0.7.0-multisource-scale-local.md).
-
-## Build and test
-
-Install development dependencies and run the tests:
-
-```powershell
-python -m pip install -e ".[dev]"
-python -m pytest -q
-```
-
-Create or resume the attribution-complete API snapshot. Legacy crawl files are
-used only to discover candidate question IDs; their HTML text is never copied
-into the public rows. Every response is checksummed and cached per request.
-
-```powershell
-python -m src.community_qa_lab.cli data api-fetch `
-  --config configs/dataset/stackexchange_api_v0.4.0.yaml `
-  --base-release data/processed/v0.2.0 `
-  --seed-root ai.stackexchange.com=../dataset/scraped_ai `
-  --seed-root datascience.stackexchange.com=../dataset/scraped_ds `
-  --seed-root stats.stackexchange.com=../dataset/scraped_cv `
-  --seed-root stats.stackexchange.com=../dataset/expanded_v5/stats `
-  --cache-dir data/cache/v0.4.0
-
-python -m src.community_qa_lab.cli data api-build `
-  --config configs/dataset/stackexchange_api_v0.4.0.yaml `
-  --cache-dir data/cache/v0.4.0 `
-  --output-dir data/processed/v0.4.0
-
-python -m src.community_qa_lab.cli data verify `
-  --output-dir data/processed/v0.4.0
-```
-
-An API key is optional. If needed, set `STACKEXCHANGE_KEY` in the process
-environment; the key is never written to a cache or manifest. An interrupted
-fetch resumes from already-validated request files.
-
-Run a CPU baseline directly on the flat Parquet split:
-
-```powershell
-python -m src.community_qa_lab.cli benchmark run `
-  --input data/processed/v0.4.0/test.parquet `
-  --ranker bm25 `
-  --output artifacts/bm25_test.jsonl
-
-python -m src.community_qa_lab.cli benchmark evaluate `
-  --input data/processed/v0.4.0/test.parquet `
-  --predictions artifacts/bm25_test.jsonl `
-  --output artifacts/bm25_test_report.json
-```
-
-## Repository layout
-
-```text
-src/community_qa_lab/   release builder, ranking and evaluation code
-apps/qa_demo/           optional FastAPI and Next.js demonstration app
-configs/                versioned dataset and benchmark configuration
-tests/                  unit tests and a small official-format fixture
-data/raw/               ignored local source dumps
-data/processed/         ignored generated release artifacts
-data/cache/             ignored resumable API response snapshots
-benchmarks/             versioned predictions, metrics and run manifests
-reports/                deterministic release-quality profiles
-docs/                   dataset, benchmark and licensing documentation
-```
-
-The builder archives the exact v0.4.0 API response snapshot as a compressed
-sidecar. Rich source records are provenance artifacts; the viewer-facing
-dataset still exposes only the stable 11-column table.
-
-## Technical QA config v0.5.0
-
-v0.5.0 deliberately does not mix general technical communities into the
-ML-focused `default`. It adds the separate `technical_qa` config using crawler
-checkpoints only as a question-ID inventory; every published field comes from
-the archived Stack Exchange API v2.3 responses.
-
-```powershell
-python -m src.community_qa_lab.cli data api-fetch `
-  --config configs/dataset/technical_qa_v0.5.0.yaml `
-  --seed-root softwareengineering.stackexchange.com=../dataset/expanded_v5/softwareengineering_stackexchange_com `
-  --seed-root scicomp.stackexchange.com=../dataset/expanded_v5/scicomp_stackexchange_com `
-  --seed-root quantumcomputing.stackexchange.com=../dataset/expanded_v5/quantumcomputing_stackexchange_com `
-  --seed-root codereview.stackexchange.com=../dataset/expanded_v5/codereview_stackexchange_com `
-  --seed-root cstheory.stackexchange.com=../dataset/expanded_v5/cstheory_stackexchange_com `
-  --cache-dir data/raw/api/v0.5.0-technical-qa
-
-python -m src.community_qa_lab.cli data api-build `
-  --config configs/dataset/technical_qa_v0.5.0.yaml `
-  --cache-dir data/raw/api/v0.5.0-technical-qa `
-  --output-dir data/processed/v0.5.0/technical_qa
-```
-
-The released config has 1,286/252/342 question groups and 3,294/641/989
-answer rows in train/validation/test. On test, answer length reaches
-Accepted@1 `0.5526`, BM25 `0.5205`, BGE-small `0.4357`, and the cross-encoder
-`0.3743`. The semantic models significantly trail BM25 on accepted-answer
-ranking under paired bootstrap, making domain shift and weak-label bias part
-of the benchmark finding. See `docs/releases/v0.5.0.md` for exact hashes,
-site counts, commands and limitations.
-
-## Scope and responsibility
-
-The current release covers `ai.stackexchange.com`,
-`datascience.stackexchange.com`, and `stats.stackexchange.com`. Source rows
-retain their URL, author and per-record license. The seed list is intentionally
-documented as seed-based rather than an exhaustive crawl. See
-`docs/dataset_card.md` for the complete schema, provenance, limitations and
-responsible-use notes.
-
-## Base-data and benchmark release v0.4.0
-
-v0.4.0 expands the ML-focused benchmark while preserving the 11-column schema.
-It pins the two pretrained model revisions, batches CPU inference, reports
-10,000-sample question-level bootstrap intervals, and publishes paired model
-comparisons plus a deterministic dataset-quality profile.
-
-```powershell
-python scripts/build_benchmark_release.py `
-  --dataset-dir data/processed/v0.4.0 `
-  --output-dir benchmarks/answer_reranking/v0.4.0 `
-  --device cpu `
-  --benchmark-version v0.4.0
-
-python scripts/build_dataset_profile.py `
-  --dataset-dir data/processed/v0.4.0 `
-  --json-output reports/v0.4.0/dataset_quality.json `
-  --markdown-output reports/v0.4.0/dataset_quality.md
-```
-
-On test, answer length reaches Accepted@1 `0.6064`, BM25 `0.5925`, BGE-small
-`0.5647`, and the cross-encoder `0.5661`. The pretrained models do not
-significantly outperform BM25; this negative result and the strong length
-baseline are part of the benchmark finding, not hidden release failures.
-
-## Benchmark release v0.2.1
-
-`v0.2.1` adds reproducible benchmark artifacts without changing the v0.2.0
-Parquet files. It runs random, answer-length, BM25, BGE-small and a pretrained
-cross-encoder on validation and test, with per-site, disagreement and
-answer-length-spread diagnostics.
-
-```powershell
-python scripts/build_benchmark_release.py `
-  --dataset-dir data/processed/v0.2.0 `
-  --output-dir benchmarks/answer_reranking/v0.2.1 `
-  --device cpu
-```
-
-See `benchmarks/answer_reranking/v0.2.1/README.md` for scores, immutable model
-revisions, run manifests and paired bootstrap comparisons.
-
-## Audit supplement v0.3.0
-
-`v0.3.0` added a separate, evaluation-only Hugging Face configuration:
-`audit_v0_3_0`. It contains one audit row for each of the 290 candidates in
-the immutable v0.2.0 test split. It remains available for historical
-reproducibility, but it is not joinable to the current v0.4.0 `default` test.
-
-The labels come from a blinded, **model-assisted manual audit**, not from
-Stack Exchange vote or acceptance information. They provide a preferred-answer
-signal for 122 of 127 question groups; four groups have no preferred answer
-and one is marked insufficient evidence. They are non-independent of a model
-assistant and must never be used to tune a model on this public test set.
-
-```python
-from datasets import load_dataset
-
-audit = load_dataset("aluha501/community-qa-lab", "audit_v0_3_0", split="test")
-```
-
-Validate a local copy and add audit metrics to already-produced predictions:
-
-```powershell
-python -m src.community_qa_lab.cli audit validate `
-  --base-test data/processed/v0.2.0/test.parquet `
-  --audit-input audits/v0.3.0/test_answer_audit.parquet `
-  --manifest audits/v0.3.0/manifest.json
-
-python -m src.community_qa_lab.cli benchmark evaluate `
-  --input data/processed/v0.2.0/test.parquet `
-  --predictions benchmarks/answer_reranking/v0.2.1/bm25_test_predictions.jsonl `
-  --audit-input audits/v0.3.0/test_answer_audit.parquet `
-  --output artifacts/bm25_test_with_audit.json
-```
-
-Read `docs/audit_protocol_v0.3.0.md` before using these labels. The audit
-table is joinable only by `(question_id, answer_id)`; it is not a replacement
-for accepted status or vote score.
-
-## Audit quality revision v0.3.1
-
-`v0.3.1` adds `audit_v0_3_1` without replacing the frozen preferences in
-`audit_v0_3_0`. It is a transparent post-unblind quality
-revision: all 122 preferred-answer choices are identical to v0.3.0, while the
-290 candidate assessments now have candidate-specific rationales and more
-conservative factuality labels.
-
-- all 262 text-only rows use `factuality=not_verifiable`;
-- only the 17 rows with recorded external sources retain supported/partially
-  supported factuality judgments;
-- 290/290 rationales are distinct, with no legacy generated prefix;
-- no assessment field deterministically encodes `is_preferred`;
-- the five immutable v0.2.1 prediction files were re-scored without model
-  tuning or rerunning inference.
-
-```python
-from datasets import load_dataset
-
-audit = load_dataset("aluha501/community-qa-lab", "audit_v0_3_1", split="test")
-```
-
-This is not a fresh blinded or independent review. Use v0.3.1 in preference
-to v0.3.0 for candidate-level analysis, and read
-`docs/audit_protocol_v0.3.1.md` before interpreting its fields. Both audit
-configs remain tied exclusively to v0.2.0.
+---
+configs:
+  - config_name: default
+    default: true
+    data_files:
+      - split: train
+        path: train.parquet
+      - split: validation
+        path: validation.parquet
+      - split: test
+        path: test.parquet
+  - config_name: corpus_rag_chunks_multisource_v0.1
+    data_files:
+      - split: train
+        path: views/corpus_rag_chunks_multisource_v0.1/train.parquet
+      - split: validation
+        path: views/corpus_rag_chunks_multisource_v0.1/validation.parquet
+      - split: test
+        path: views/corpus_rag_chunks_multisource_v0.1/test.parquet
+  - config_name: corpus_rag_quality_multisource_v0.1
+    data_files:
+      - split: train
+        path: views/corpus_rag_quality_multisource_v0.1/train.parquet
+      - split: validation
+        path: views/corpus_rag_quality_multisource_v0.1/validation.parquet
+      - split: test
+        path: views/corpus_rag_quality_multisource_v0.1/test.parquet
+  - config_name: issue_resolution_multisource_v0.1
+    data_files:
+      - split: train
+        path: views/issue_resolution_multisource_v0.1/train.parquet
+      - split: validation
+        path: views/issue_resolution_multisource_v0.1/validation.parquet
+      - split: test
+        path: views/issue_resolution_multisource_v0.1/test.parquet
+  - config_name: preference_dpo_multisource_v0.1
+    data_files:
+      - split: train
+        path: views/preference_dpo_multisource_v0.1/train.parquet
+      - split: validation
+        path: views/preference_dpo_multisource_v0.1/validation.parquet
+      - split: test
+        path: views/preference_dpo_multisource_v0.1/test.parquet
+  - config_name: reranking_listwise_multisource_v0.1
+    data_files:
+      - split: train
+        path: views/reranking_listwise_multisource_v0.1/train.parquet
+      - split: validation
+        path: views/reranking_listwise_multisource_v0.1/validation.parquet
+      - split: test
+        path: views/reranking_listwise_multisource_v0.1/test.parquet
+  - config_name: v0.8.0_rag_quality
+    data_files:
+      - split: train
+        path: releases/v0.8.0-rag-quality/corpus_rag_chunks_multisource_v0.1/train.parquet
+      - split: validation
+        path: releases/v0.8.0-rag-quality/corpus_rag_chunks_multisource_v0.1/validation.parquet
+      - split: test
+        path: releases/v0.8.0-rag-quality/corpus_rag_chunks_multisource_v0.1/test.parquet
+license: apache-2.0
+task_categories:
+  - question-answering
+  - sentence-similarity
+  - text-generation
+language:
+  - en
+tags:
+  - rag
+  - code
+  - modern-ai
+  - dpo
+  - preference
+  - issue-resolution
+  - systems
+  - llm-engineering
+size_categories:
+  - 10K<n<100K
+---
+
+# Community QA Lab — Modern AI & Systems (v0.9.0 & v0.8.0)
+
+A high-signal, multi-task dataset collected from **24 leading open-source AI, LLM runtime, agent, and modern systems repositories** spanning **2020 through 2026**.
+
+This release transitions beyond legacy programming FAQ dumps to focus exclusively on contemporary engineering challenges: **quantization kernels, FSDP/distributed training, CUDA OOM troubleshooting, vLLM/SGLang serving, and autonomous agent loops**.
+
+---
+
+## 📊 Dataset Summary & Multi-Task Views
+
+| Task / View | Total Samples | Train | Validation | Test | Description |
+| :--- | ---: | ---: | ---: | ---: | :--- |
+| **`corpus_rag_chunks_multisource_v0.1`** *(default)* | **26,780** | 25,044 | 1,454 | 282 | Syntax-preserving, code-aware RAG chunks (avg 326 tokens, 8.7M tokens total). |
+| **`corpus_rag_quality_multisource_v0.1`** | **20,143** | 19,070 | 923 | 150 | Full post and thread document view with context & metadata. |
+| **`issue_resolution_multisource_v0.1`** | **3,600** | 3,472 | 128 | 0 | Paired Issue Problem Statements $\rightarrow$ Merged PR resolutions for Coding Agents. |
+| **`preference_dpo_multisource_v0.1`** | **1,756** | 1,727 | 18 | 11 | Direct Preference Optimization (DPO) chosen vs. rejected response pairs. |
+| **`reranking_listwise_multisource_v0.1`** | **357** | 335 | 14 | 8 | Listwise candidate answers with interaction-driven relevance scores. |
+| **`v0.8.0_rag_quality`** *(legacy archive)* | **62,159** | 41,634 | 10,265 | 10,260 | Legacy StackExchange cleaned RAG corpus view. |
+
+---
+
+## 🔍 Key Insights & Exploratory Data Analysis (EDA)
+
+* **Total Token Count:** **8,729,294 tokens** (`cl100k_base` / GPT tokenizer equivalent).
+* **Code Density:** **43.8%** of chunks contain runnable code blocks.
+* **Top Programming Languages:** Python (77.4%), Shell/Bash (9.0%), SQL (2.4%), JSON (2.1%), C++/CUDA (0.6%), Rust (0.6%).
+* **Error Taxonomy Coverage:**
+  - *Quantization & Kernels (FlashAttention, Triton, QLoRA, GGUF):* 2,178 cases
+  - *Distributed & Multi-GPU (FSDP, DeepSpeed Zero-3, NCCL):* 1,801 cases
+  - *Data & Type Mismatch (ValueError / TypeError):* 1,014 cases
+  - *Runtime & Memory (CUDA OOM, RuntimeError):* 1,675 cases
+
+---
+
+## 🏛️ Repositories Included (24 Top Tier Projects)
+
+- **LLM Training & Fine-tuning:** `huggingface/transformers`, `pytorch/pytorch`, `huggingface/trl`, `unslothai/unsloth`, `huggingface/peft`, `huggingface/accelerate`, `huggingface/datasets`, `deepspeedai/DeepSpeed`, `ray-project/ray`
+- **Inference & Serving:** `vllm-project/vllm`, `sgl-project/sglang`, `ollama/ollama`, `ggerganov/llama.cpp`
+- **Agents & RAG Orchestration:** `run-llama/llama_index`, `langchain-ai/langchain`, `langchain-ai/langgraph`, `crewAIInc/crewAI`, `huggingface/smolagents`
+- **Modern Infrastructure & Tools:** `astral-sh/uv`, `pydantic/pydantic`, `tiangolo/fastapi`, `pola-rs/polars`, `duckdb/duckdb`, `gradio-app/gradio`
+
+---
+
+## 🚀 Quickstart & Usage
+
+### 1. Load Clean RAG Chunks (Default View)
+```python
+from datasets import load_dataset
+
+# Load default code-aware chunks (v0.9.0)
+dataset = load_dataset("aluha501/community-qa-lab", split="train")
+
+print(dataset[0])
+# {
+#   "chunk_id": "gh:tiangolo/fastapi:disc:1234:chunk:0",
+#   "chunk_text": "...",
+#   "has_code": True,
+#   "code_languages": ["python"],
+#   "repository": "tiangolo/fastapi",
+#   "source_type": "github_discussion"
+# }
+```
+
+### 2. Load Coding Agent Issue-Resolution Pairs
+```python
+issue_dataset = load_dataset(
+    "aluha501/community-qa-lab",
+    "issue_resolution_multisource_v0.1",
+    split="train"
+)
+```
+
+### 3. Load DPO Preference Pairs for Alignment
+```python
+dpo_dataset = load_dataset(
+    "aluha501/community-qa-lab",
+    "preference_dpo_multisource_v0.1",
+    split="train"
+)
+```
+
+### 4. Load Legacy v0.8.0 Clean RAG Quality View
+```python
+v8_dataset = load_dataset(
+    "aluha501/community-qa-lab",
+    "v0.8.0_rag_quality",
+    split="train"
+)
+```
+
+---
+
+## 🛡️ License & Provenance
+
+All collected data originates from official public repositories under **permissive open-source licenses** (Apache-2.0, MIT, BSD-3-Clause). Private tokens, API keys, and personal emails have been sanitized.

releases/v0.8.0-rag-quality/corpus_rag_chunks_multisource_v0.1/test.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a65adb33a972930898eb21db971b4708f2bcbc980f5176049f9b4740bbd2524e
+size 8919350

releases/v0.8.0-rag-quality/corpus_rag_chunks_multisource_v0.1/train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d880e6bcf1f87e81a185f4f83b2cbed683fc6d3f470ed90f10fdac362135c42a
+size 23347101

releases/v0.8.0-rag-quality/corpus_rag_chunks_multisource_v0.1/validation.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b23cf2949432c1b3fcbc9b403d43ef28b26d63d9fead5c19af195d2c925da598
+size 7324567

releases/v0.8.0-rag-quality/corpus_rag_quality_multisource_v0.1/test.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d91664ed0f548adc15c785e373bf133b9a184191eb4650f5af177ed3b02f7393
+size 16616282

releases/v0.8.0-rag-quality/corpus_rag_quality_multisource_v0.1/train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e5e09763a2206dd7a8f45a0f343447154dafe5525d5300a1fda9cbbdcaa54b86
+size 42756115

releases/v0.8.0-rag-quality/corpus_rag_quality_multisource_v0.1/validation.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1273a8ec2dffc39a5383a1c6c7fc132ea3ef69d6f8f380c1f6f9f716e1aa1d38
+size 13668563

test.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c611508cac16b0028161539f4c85456e029e7e5bd16582147d5c6f445e6a62b2
-size 12696133
+oid sha256:be5818f847fda202b7b1aba373a54d10e33c5c24002d34ef73af201017cf93d9
+size 128411

train.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f39c152470645d20b44bc780610cfaa41e720ba19c51f62bebee9cd04bfb7f50
-size 32881836
+oid sha256:82eab8cecd417ae54e9e8fd36c7b485f9ec938801e9101d433c4bce3e935f6a0
+size 12526458

validation.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4eedcf05628e6cbb27259eafef83d7961131b371063736f97cb05efd3fe814cb
-size 10346965
+oid sha256:ed1b8603536e437218de4d5035103cc66cec00091bfe413da248035a13708f37
+size 558289

views/corpus_rag_chunks_multisource_v0.1/test.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be5818f847fda202b7b1aba373a54d10e33c5c24002d34ef73af201017cf93d9
+size 128411

views/corpus_rag_chunks_multisource_v0.1/train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:82eab8cecd417ae54e9e8fd36c7b485f9ec938801e9101d433c4bce3e935f6a0
+size 12526458

views/corpus_rag_chunks_multisource_v0.1/validation.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed1b8603536e437218de4d5035103cc66cec00091bfe413da248035a13708f37
+size 558289

views/corpus_rag_quality_multisource_v0.1/test.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2535b08afa081eefb4d686b619da9542b21f0ce4f34428549885d8e683493cc0
+size 219713

views/corpus_rag_quality_multisource_v0.1/train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ec979ca408a78da8285466de46fa0b03e619858127bfbbd7e98d7aad8ac38e8
+size 22620748

views/corpus_rag_quality_multisource_v0.1/validation.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f2d464fc2a6588ef597ab71c30bfd3baf2497572bc01a3ecd3a37f6779b263e
+size 913934

views/issue_resolution_multisource_v0.1/test.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:27388fd937ebf0072634cae82a3219b823a17698d566cc2ad190623a75878337
+size 3697

views/issue_resolution_multisource_v0.1/train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed2f48069b3ceaf6b5e9b59ff1a9d9aa40a147b1ad85a05141fb43dfcdf41797
+size 1098208

views/issue_resolution_multisource_v0.1/validation.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60f75061978b268567afe0ede858c7ce7e9b285f931aef584d635c87f15c462f
+size 60228

views/preference_dpo_multisource_v0.1/test.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24499b4df1d076ffa22c8b4b216c737a18d2538b565e0d5b39edc77cacb47255
+size 24071

views/preference_dpo_multisource_v0.1/train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d32254bea109979b92fae06b6c0aa2a6071ce95b1ffb6118fd8c5e5732e9e7aa
+size 653129

views/preference_dpo_multisource_v0.1/validation.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf7755c278524d5b119f3d2e47be6ceb34cc7ae1e988aa715a0a69d2a4500fe3
+size 27526

views/reranking_listwise_multisource_v0.1/test.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:85c51ee9ab4824954ab060657b0c356f0d11b03e5a73e2d170851e4c988b45ae
+size 18028

views/reranking_listwise_multisource_v0.1/train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d66650fe12ad62ef58c5c980602dfd38654815aeffa0f9cae071508a8facf136
+size 563987

views/reranking_listwise_multisource_v0.1/validation.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d67ec677b6c1683b930fd8cd97c6cd257477767dfb50f1235e4510ebc8b642c
+size 27061