sarthak20024/sih26099-cpse-material-codes

SIH 26099 — Collected Dataset

AI-Driven Standardization & Harmonization of Material Codes Across CPSEs

This workspace holds the data-collection stage only — no model, no training, no feature engineering. Just raw public sources, their extracted structured form, and the reference taxonomies/vocabularies the harmonisation step needs.

Collected live on 2026-09-08. All row counts below were verified by reading the files back with pandas.

---

1. Headline numbers

Dataset	Rows	File
Material description corpus (main table)	21,513	data/processed/extracted_items/material_description_corpus.csv
NTPC NIT metadata	1,418	data/raw/ntpc/tenders/nit_details.csv
NTPC document index	1,626	data/raw/ntpc/tenders/documents_index.csv
NTPC per-document text	1,615	data/processed/extracted_items/ntpc_doc_text.csv + ntpc_text/
NTPC tables pulled from PDFs	78,216	data/processed/extracted_items/ntpc_doc_tables.csv
NTPC structured material lines	486	data/processed/extracted_items/ntpc_material_items.csv
IOCL tender metadata	179	data/raw/iocl/tenders/tender_details.csv
IOCL bid-cover inventory	3,715	data/raw/iocl/tenders/bid_covers.csv
IOCL procurement-plan items	1,224	data/processed/extracted_items/iocl_procurement_plan_items.csv
OIL India live tenders	180 (140 unique)	data/raw/oil_india/tenders/live_tenders.csv
OIL India archived tenders	4,051 (4,016 unique)	data/raw/oil_india/tenders/archive_tenders.csv
OIL India historical tenders	12,116 (11,987 unique)	data/raw/oil_india/tenders/old_tenders.csv
OIL India tender detail pages	4,165 (4,152 unique)	data/raw/oil_india/tenders/tender_details.csv
OIL India NIT download links	4,152	data/raw/oil_india/tenders/nit_documents_index.csv
OIL India corrigenda	145	data/raw/oil_india/tenders/corrigenda.csv
UNSPSC master taxonomy	71,502	data/reference/unspsc/unspsc_master.csv
UNSPSC industrial subset	12,771	data/reference/unspsc/unspsc_industrial_subset.csv
CPPP organisations + live counts	235	data/raw/cppp/tenders/organisations_with_tender_counts.csv

---

2. What was reachable, and what was not

Source	Status	What we got
NTPC ntpctender.ntpc.co.in	✅ full	1,418 NITs, 1,626 documents indexed, 1,615 PDFs text-extracted
IOCL iocletenders.nic.in	✅ full	179 tenders, full metadata, 3,715 bid-cover rows
IOCL iocl.com	✅ full	104-page Future Procurement Plan → 1,224 item rows + 2 item-list PDFs
OIL India oil-india.com	✅ full	180 live, 4,016 archived, 11,987 historical tenders, 145 corrigenda, EoI/pre-bid, 4,152 NIT links
UNSPSC	✅ full	71,502 commodities / 5,313 classes / 465 families / 57 segments
CPPP eprocure.gov.in	⚠️ partial	Portal reachable, but every search POST is captcha + session-gated. Only the org list (235 orgs / 2,452 live tenders) and the 85-category taxonomy came through
SAIL eproc.sail.co.in	❌ blocked	REST API (/rest/openarea/*) discovered, but every call returns a generic error page (WAF)
ONGC	❌ unreachable	ongcindia.com / ongc.co.in hosts time out from this network
BPCL	❌ blocked	HTTP 403 (CloudFront)
Coal India	❌ blocked	HTTP 403
GeM gem.gov.in	❌ unreachable	Host times out — but GeM data reaches us indirectly (see below)
EIL tenders.eil.co.in	❌ unreachable	Host times out
data.gov.in	❌ gated	Public API needs a key; the catalogue site is client-side rendered
OIL India SAP SRM	❌ login-only	etender.srm.oilindia.in serves a logon page

GeM is still well represented, just not from the GeM portal: OIL India has a whole GeM Tenders category (2,820 archived + 102 live records), 2,085 historical tenders are categorised GEM, and NTPC NITs carry GeM item codes and GeMARPTS search strings.

---

3. Directory map

data/
├── raw/                                     untouched source output
│   ├── ntpc/tenders/     nit_details.csv, listing_rows.csv,
│   │                     documents_index.csv, raw_html_sample/ (60)
│   ├── ntpc/documents/   70 sample NIT/corrigendum PDFs (full index kept)
│   ├── iocl/tenders/     tender_details.csv, search_rows.csv, bid_covers.csv,
│   │                     3 source PDFs, raw_html_sample/ (50)
│   ├── oil_india/tenders/  live/archive/old/tender_details/nit_documents_index/
│   │                       corrigenda/eoi CSVs, raw_html_sample/ (60)
│   ├── oil_india/documents/  sample NIT PDF
│   └── cppp/tenders/     organisations_with_tender_counts.csv
├── reference/
│   ├── unspsc/           unspsc_master.csv (71,502),
│   │                     unspsc_industrial_subset.csv (12,771),
│   │                     oklahoma_unspsc.csv (original download)
│   ├── abbreviations/    material_abbreviations.csv (154),
│   │                     material_grades.csv (58), pressure_classes.csv (28)
│   ├── units/            unit_normalisation.csv (125)
│   ├── cppp_product_categories.csv (85 categories)
│   └── ntpc_locations.csv (116 units/projects)
└── processed/extracted_items/
    ├── material_description_corpus.csv      ← main working table
    ├── ntpc_doc_text.csv, ntpc_doc_tables.csv
    ├── ntpc_material_items.csv
    ├── iocl_procurement_plan_items.csv
    └── ntpc_text/                           1,615 per-document .txt

Total 119 MB, 1,894 files. PRUNED.txt lists the 7,133 bulky files (1.3 GB) that were removed to stay inside the snapshot budget — every one is re-downloadable from the URLs kept in the CSVs, and the harvest scripts reproduce them.

---

4. The main table: material_description_corpus.csv

One row = one real, publicly published material/item description, tagged with the organisation that published it and where it came from.

column	meaning
corpus_id	stable row id (M000001…)
organization	NTPC / IOCL / Oil India
source_system	host the record came from
source_section	which listing, page or document
tender_reference, tender_id	publisher's own identifiers
description	the material/item description text
description_kind	tender_title, nit_brief_description, historical_tender, tender_detail, procurement_plan_item, doc_spec_string, …
item_type_hint	coarse type from a keyword pass (valve, pump, pipe, cable…)
quantity, unit	where the source gave them
location, product_category	publisher metadata
document_url, source_url	provenance — always kept

Split by organisation: Oil India 18,950 · NTPC 1,843 · IOCL 720. 17,400 of the 21,513 descriptions are distinct strings; the rest repeat across organisations or across a tender's title / work-description pair.

Cross-organisation near-duplicates are kept on purpose — they are exactly the variation the harmonisation step has to resolve. Only exact duplicates within one organisation and description kind are collapsed.

Coarse type distribution (top): pipe 1,214 · pump 927 · tank 580 · cable 330 · steel 325 · instrument 259 · spares 246 · chemical 236 · valve 224 · pipe fitting 200 · compressor 184 · safety valve 183 · transformer 151 · flange 141 · gate valve 131.

---

5. Notable content

Real GeM catalogue strings from NTPC NIT PDFs — already semi-structured material codes:

CABLE,PWR,240MM2,1C,STRANDED,AL,11KV
CABLE, PWR, 150MM2, 1C, STRANDED, AL, 11KV
GASKET,17.5KV,1000A,AREVA
GASKET,3.6KV,2000A,AREVA
SEAMLESS,SS,A312-TP304L,80S,15MM , PIPE
PIPE:BLK, CORTEN,A423GR1,ERW,2MM,40MM

IOCL procurement-plan items with quantities, units and values:

Vessels/Drums/Tanks                          99 Nos.   ₹108.06 Cr
Pumps                                       275 Nos.   ₹138.08 Cr
Heat Ex/Air coolers/Fin Fan Coolers         130 Nos.   ₹142.66 Cr
Rotary Valves/Safety Valves/Control Valves/On-off Valves…
Synthetic base oil Poly alfa olefin 4 CST, PAO 40(ST 55578)
Benzene hydrogenation catalyst for ISOM unit
Sample Coolers (PI) & Sample Bombs (PE)

OIL India historical tenders — the densest source of true material text:

Line Pipe, 2" , ERW Galvanised, Screwed (Q3)
Conventional Gas Lift Mandrels
PMCC PANEL – 01 NOS.
OXYGEN GAS STORAGE CYLINDER, 7 CUM (Q3), ACETYLENE GAS STORAGE CYLINDER, 6-7 CUM (Q3)
GATE VALVE -STEEL
High pressure high temperature (HPHT), Four cells (175ml), filter press (Q3)
SUPPLY OF CHAIN TONGS AND PIPE WRENCH UNDER RATE CONTRACT
Supply of Stage cementing collars

IOCL tender titles — long, specification-rich free text, exactly what needs normalising:

Procurement of Capital Spare, Complete Trip and Throttle valve (Make Siemens
Energy Industrial Turbomachinery India Private Limited) of WGC Turbine as per
Insurance spares guideline rotary 2025 26 at IOCL Paradip Refinery

SP/B269-000-WB-MR-3741/396 - VALVES-GATE, GLOBE, CHECK FOR P-25 PROJECT

---

6. Reference tables

file	rows	purpose
unspsc_master.csv	71,502	full UNSPSC segment→family→class→commodity hierarchy — the target code system
unspsc_industrial_subset.csv	12,771	the 20 industrial segments relevant to CPSE materials
material_abbreviations.csv	154	BALL VL→BALL VALVE, NRV→CHECK VALVE, RF→RAISED FACE, …
material_grades.csv	58	ASTM / API / IS grade tokens → standard, grade, description
pressure_classes.csv	28	300# / CL 300 / CLASS-300 → CLASS 300 (≈PN50)
unit_normalisation.csv	125	EA/NOS/PCS→NOS, inch↔mm size equivalences, mass/volume/pressure factors
cppp_product_categories.csv	85	CPPP's own product-category taxonomy
ntpc_locations.csv	116	NTPC stations / projects

---

7. Scripts

script	what it does
scripts/harvest_ntpc.py	NTPC listing sweep (types + 130 material keywords), NIT detail pages, corrigendum JSON API, document download
scripts/harvest_iocl.py	IOCL captcha-free keyword sweep; search and detail fetch share one session (required)
scripts/harvest_oil_india.py	OIL India live / rest / old / details subcommands, checkpointed
scripts/extract_ntpc_docs.py	memory-safe text + table extraction over the NTPC PDFs
scripts/extract_ntpc_material_items.py	GeM item strings, BOQ lines, spec strings from the NIT text
scripts/extract_iocl_plan.py	IOCL Future Procurement Plan → item rows
scripts/extract_oil_nit_links.py	recover NIT download links from saved detail pages
scripts/build_reference_tables.py	abbreviation / unit / grade / pressure-class tables
scripts/build_material_corpus.py	merge everything into material_description_corpus.csv
scripts/prune_workspace.py	trim reproducible bulk, log to PRUNED.txt

Reproduce everything with:

python3 scripts/harvest_ntpc.py
python3 scripts/harvest_iocl.py            # IOCL_GENERIC=1 for the wider sweep
python3 scripts/harvest_oil_india.py live
python3 scripts/harvest_oil_india.py rest
python3 scripts/harvest_oil_india.py old
python3 scripts/harvest_oil_india.py details
python3 scripts/extract_ntpc_docs.py
python3 scripts/extract_ntpc_material_items.py
python3 scripts/extract_iocl_plan.py
python3 scripts/extract_oil_nit_links.py
python3 scripts/build_reference_tables.py
python3 scripts/build_material_corpus.py

---

8. Four things worth knowing about these portals

1. IOCL's detail URLs are session-bound. The listing hands out …&sp=<token> links that only resolve for a session which has already fetched the Home page. A fresh session gets a 1,410-byte stub with zero fields instead of the 79,870-byte page. harvest_iocl.py therefore searches and fetches inside one session.

2. IOCL's search caps at 20 hits per keyword, most-recent first, with no pagination. Broad terms (the, supply, procurement) each return a different 20, so the harvester sweeps 230 keywords and unions the results — 179 distinct tenders.

3. OIL India's Drupal pager loops past the end of a result set. The archive listing advertises data-total=450 but keeps returning rows forever: a naive crawl produced 19,570 rows holding only 449 distinct tender numbers (97.6 % duplicates). The harvester now stops when a page yields no unseen tender number. It also has to tolerate transient blank pages — page 378 of /old-tender comes back empty while pages 379–1,311 all carry data, so an early version gave up 9,000 rows short.

4. OIL India's NIT links have no file extension. They are served by /download-tender-document?ten_detail=… and answer binary/octet-stream, so matching on .pdf finds nothing and the extension has to come from the payload magic bytes. All 4,152 links are indexed in nit_documents_index.csv; only a sample is downloaded because the advertised sizes run to several GB in total.