Update README.md

README.md

@@ -1,6 +1,6 @@
 ---
 license: cc-by-sa-4.0
-pretty_name: NFCorpus (corpus.tsv re-encoding)
+pretty_name: NFCorpus (corpus.jsonl re-encoding)
 task_categories:
   - sentence-similarity
   - text-retrieval
@@ -21,11 +21,11 @@ source_datasets:
   - original
 ---
 
-# NFCorpus — `corpus.tsv` re-encoding
+# NFCorpus — `corpus.jsonl` re-encoding
 
-A re-encoding of the [NFCorpus](https://www.cl.uni-heidelberg.de/statnlpgroup/nfcorpus/) medical-IR corpus by Boteva et al. (Heidelberg StatNLP, ECIR 2016), distilled into a single gzip-compressed tab-delimited file. The 3,633 rows, integer `passage_id` column, and concatenated `text` column derive verbatim from the BEIR-reformatted [`corpus.jsonl`](https://public.ukp.informatik.tu-darmstadt.de/thakur/BEIR/datasets/nfcorpus.zip) — only the outer container differs.
+A re-encoding of the [NFCorpus](https://www.cl.uni-heidelberg.de/statnlpgroup/nfcorpus/) medical-IR corpus by Boteva et al. (Heidelberg StatNLP, ECIR 2016), distilled into a single gzip-compressed JSON-Lines file. The 3,633 rows and the four-key per-line schema (`_id`, `title`, `text`, `metadata`) match the [BEIR-reformatted](https://public.ukp.informatik.tu-darmstadt.de/thakur/BEIR/datasets/nfcorpus.zip) `corpus.jsonl` byte-for-byte — only the outer container differs.
 
-Re-hosted under Heliosoph for ingestion-pipeline stability — the upstream BEIR archive (`nfcorpus.zip`) wraps the corpus alongside queries and qrels in a zip, and the Heidelberg release splits docs across three per-split TSVs with overlap. Dropping the zip wrapper and unifying to one deduped TSV lets a streaming gzip CSV reader work directly against the file.
+Re-hosted under Heliosoph for ingestion-pipeline stability — the upstream BEIR archive (`nfcorpus.zip`) wraps `corpus.jsonl` alongside queries and qrels in a single zip, and the original Heidelberg release splits docs across three per-split TSVs with overlap. Dropping the zip wrapper down to a plain `corpus.jsonl.gz` lets a streaming JSON-Lines reader work directly against the file with no intermediate extraction step.
 
 Credit: Vera Boteva, Demian Gholipour, Artem Sokolov, Stefan Riezler (Heidelberg StatNLP, ECIR 2016) — corpus extracted from [NutritionFacts.org](https://nutritionfacts.org/) under CC BY-SA 4.0. BEIR-reformatted distribution by Nandan Thakur et al. (UKP, TU Darmstadt, NeurIPS 2021).
 
@@ -34,71 +34,68 @@ Credit: Vera Boteva, Demian Gholipour, Artem Sokolov, Stefan Riezler (Heidelberg
 The two upstream NFCorpus distributions both add work for an ingestion pipeline that just wants the passages:
 
 - **Heidelberg's `nfcorpus.tar.gz`** splits documents across `train.docs`, `dev.docs`, and `test.docs` with overlap — recovering the canonical 3,633 unique passages requires a union+dedup step on the doc-id column.
-- **BEIR's `nfcorpus.zip`** wraps the unified `corpus.jsonl` together with queries and qrels in a zip, and the `corpus.jsonl` schema (`{"_id": ..., "title": ..., "text": ..., "metadata": {}}`) needs a JSON-line reader to flatten.
+- **BEIR's `nfcorpus.zip`** wraps the unified `corpus.jsonl` together with queries and qrels in a zip — pipelines that ingest one corpus file at a time have to either extract the zip first or know which member to seek to.
 
-This mirror strips both wrappers down to a single gzipped TSV — `corpus.tsv.gz` — that any streaming CSV reader can consume directly. Title and body text are concatenated as `title + ". " + text` (the convention used by most BEIR embedder evaluations) so the file fits the canonical two-column passage-retrieval shape.
+This mirror strips the zip wrapper down to a single gzipped JSON-Lines file — `corpus.jsonl.gz` — that any streaming JSONL reader can consume directly. The line-level schema is preserved verbatim, so the `_id` field can still be joined against the official BEIR `qrels/test.tsv` for nDCG@10 evaluation without translation.
 
-The decompressed `_id` and concatenated `title + text` content matches BEIR's `corpus.jsonl` exactly for all 3,633 rows.
+The decompressed contents match BEIR's `corpus.jsonl` line-for-line, byte-for-byte.
 
 ## What this repo contains
 
 ```
-corpus.tsv.gz     # 3,633 rows, ~1.8 MB compressed, UTF-8, tab-delimited, no header
+corpus.jsonl.gz   # 3,633 lines, ~1.9 MB compressed, UTF-8, one JSON object per line
 ```
 
-One file, two columns, **no header row**:
+One file, one JSON object per non-empty line, four keys per object:
 
-| Column | Type | Meaning |
+| Key | Type | Meaning |
 |---|---|---|
-| 0 | int | `passage_id` — stable id from the upstream BEIR `_id` field. |
-| 1 | string | `text` — `title + ". " + body`, original casing + punctuation. |
+| `_id` | string | BEIR-style passage id (e.g. `"MED-10"`). Stable join key against the official `queries.jsonl` and `qrels/test.tsv`. |
+| `title` | string | Article title from NutritionFacts.org, original casing + punctuation. |
+| `text` | string | Article body, original casing + punctuation. |
+| `metadata` | object | `{"url": "..."}` — the source NutritionFacts.org permalink. |
 
-The `passage_id` is the join key used by the official NFCorpus / BEIR `qrels` (relevance judgments) — if you intend to evaluate against the standard test split or compare against MTEB / BEIR leaderboards, keep the id intact end-to-end.
+Lines are not sorted by `_id` in any meaningful order; the order matches BEIR's upstream `corpus.jsonl` exactly.
 
 ## How to use
 
-Stream the gzipped TSV directly with pandas:
+Stream the gzipped JSONL row-by-row with the standard library (zero-memory streaming, ~3,633 dict allocations):
 
 ```python
-import pandas as pd
-
-df = pd.read_csv(
-    "corpus.tsv.gz",
-    sep="\t",
-    header=None,
-    names=["passage_id", "text"],
-    dtype={"passage_id": "string", "text": "string"},
-    compression="gzip",
-)
-print(df.shape)  # (3633, 2)
-print(df["text"].str.len().describe())
+import gzip, json
+
+with gzip.open("corpus.jsonl.gz", "rt", encoding="utf-8") as f:
+    for line in f:
+        rec = json.loads(line)
+        passage_id = rec["_id"]
+        title      = rec["title"]
+        body       = rec["text"]
+        # Most BEIR embedder evaluations concatenate title + body before encoding:
+        passage    = f"{title}. {body}"
 ```
 
-Or row-by-row with the standard library (zero-memory streaming):
+Or load the whole corpus into a pandas frame (fits in <50 MB RAM):
 
 ```python
-import csv, gzip
+import pandas as pd
 
-with gzip.open("corpus.tsv.gz", "rt", encoding="utf-8", newline="") as f:
-    for pid, text in csv.reader(f, delimiter="\t"):
-        # pid: BEIR-style string id (e.g. "MED-10")
-        # text: "<title>. <body>" — full passage, ready to embed
-        ...
+df = pd.read_json("corpus.jsonl.gz", lines=True, compression="gzip")
+print(df.shape)              # (3633, 4)
+print(df.columns.tolist())   # ['_id', 'title', 'text', 'metadata']
+print(df["text"].str.len().describe())
 ```
 
 ## Dataset specs
 
 | | Spec |
 |---|---|
-| Rows | 3,633 passages |
-| Distinct passages | 3,633 (passage_id is unique) |
+| Lines | 3,633 passages |
+| Distinct passages | 3,633 (`_id` is unique) |
 | Encoding | UTF-8 |
-| Delimiter | Tab (`\t`) |
-| Header | None — first row is data |
-| Quoting | Minimal — embedded tabs in passage text are rare; standard CSV/TSV readers handle them with double-quote escaping |
-| Compressed size | ~1.8 MB |
-| Uncompressed size | ~5.93 MB |
-| Average passage length | ~232 words / ~330 BERT word-piece tokens |
+| Format | JSON Lines (one JSON object per non-empty line, no enclosing array) |
+| Compressed size | ~1.9 MB |
+| Uncompressed size | ~5.9 MB |
+| Average passage length | ~232 words / ~330 BERT word-piece tokens (title + body concatenated) |
 | Max passage length | ~5,000+ tokens — exercises long-context embedders |
 | Language | English |
 | Domain | Medical / nutrition (NutritionFacts.org articles) |
@@ -110,7 +107,7 @@ with gzip.open("corpus.tsv.gz", "rt", encoding="utf-8", newline="") as f:
 - **Long-context embedder showcase**: the passage length tail runs into the thousands of tokens — full articles, not snippets. Embedders with 4K / 8K context (Jina v2, E5 long, BGE long) actually see passages worth more than the first 512 tokens.
 - **BEIR / MTEB leaderboard comparison**: NFCorpus is one of the original BEIR zero-shot retrieval benchmarks, so a fresh nDCG@10 number from your recipe is directly comparable to dozens of published embedders.
 
-For the much larger general-domain counterpart, reach for **MS MARCO Passages** — same two-column shape, 2,400× the rows, web-search vocabulary instead of medical.
+For the much larger general-domain counterpart, reach for **MS MARCO Passages** — passage-retrieval shape, 2,400× the rows, web-search vocabulary instead of medical.
 
 ## Companion splits (not in this repo)
 
@@ -119,7 +116,7 @@ A full NFCorpus / BEIR retrieval evaluation requires two more files from the ups
 - `queries.jsonl` — 3,237 natural-language health queries (`_id`, `text`).
 - `qrels/test.tsv` — graded (query_id, passage_id, relevance) judgments for the test split.
 
-Run a full nDCG@10 / MRR@10 evaluation by pairing this corpus with those two. A follow-up Heliosoph variant may bundle them as `queries.tsv.gz` + `qrels_test.tsv.gz`.
+Run a full nDCG@10 / MRR@10 evaluation by pairing this corpus with those two. A follow-up Heliosoph variant may bundle them.
 
 ## License