add OCR markdown of the PDF corpus as parquet shards

markdown/README.md

@@ -0,0 +1,28 @@
+# markdown/
+
+Full text of the PDFs under `pdfs/`, produced with `deepseek-ai/DeepSeek-OCR-2`
+(vLLM 0.28.0, one page per request, gundam crop mode).
+
+One row per paper:
+
+| column | meaning |
+| --- | --- |
+| `id` | stable id derived from the PDF path |
+| `venue`, `year` | parsed from `pdfs/<venue>/<year>/` |
+| `pdf_path` | path of the source PDF inside this repo |
+| `pages` | pages sent to the model |
+| `markdown` | the text, with `<\|ref\|>` unwrapped and `<\|det\|>` boxes dropped |
+| `n_chars` | length of `markdown` |
+| `degenerate_pages` | pages the OCR pass flagged as looping or as a collapsed table |
+| `ocr_seconds` | wall time for the paper |
+
+```python
+from datasets import load_dataset
+ds = load_dataset("JonesLin/writing-model-papers-2016-2021",
+                  data_files="markdown/*.parquet", split="train")
+```
+
+Quality: on an 80-paper sample, 52% were clean, 32% had mild repetition and 17% had a
+badly degenerated page. Extracting the same PDFs with PyMuPDF puts 8% in that last bucket,
+so part of the gap is repetition that is genuinely in the papers. Filter on
+`degenerate_pages == 0` for the strictest subset.

markdown/train-00000.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3d8cb2706df9a9ea750ab94aff2e27eff57283082ce347a652df6447f973d7f9
+size 67048114

markdown/train-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:858182a39f34c6cbc3f467e6634c7243c9c31e32e7b1985044ff22fe6314d133
+size 67186558

markdown/train-00002.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5fb215053f738d2b60d30984b05e06c05a63b6b12698a6ab635ac2f08e0e6b0
+size 67473318

markdown/train-00003.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a9ec62ff77f53c37e4bf935c888ea92e042ef1f57641e718bcad28971ea5180e
+size 67601648

markdown/train-00004.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3a522f96133c88f0b7c5dd20e978fe8892f2303f627a98941a90c6d93209a5b6
+size 68941821

markdown/train-00005.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f2b57f66d5741016a69360fe4eadeb669067f7e9db7b3786add4105837a90f50
+size 68670078

markdown/train-00006.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd703f81387fa0d36e062070ef450233a9e89f1270cd98217ebab400c6651741
+size 67490184

markdown/train-00007.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd52c10a7e55d3050fb3d7749c4d46cfe724349da0f249e25982c9142b04b1d5
+size 67384096

markdown/train-00008.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e30d9b3a528c0fcf11dd2b747f360135347d7cb14a6a1c4326e18b8d9923fad3
+size 67746319

markdown/train-00009.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:84b0b817f828b9727c27b945d5a160284903da30f75f3cd01140a893280450db
+size 9922508