SOB-FT-Extract

Unified fine-tuning corpus for JSON extraction and tool calling, built by mapping upstream datasets onto a shared schema and merging them here.

Each upstream corpus is also published standalone under mariem123kfg/sob-ft-* with its own dataset card documenting source-specific changes.

Merged sources (118,799 rows)

Filter any subset with source_dataset:

Rows	source_dataset	Standalone repo	Task
35,025	scrapegraphai_full	—	Grounded JSON extraction (web scrape)
19,257	boradorish/STAGE-eval	sob-ft-stage-eval	Grounded JSON extraction (long reports)
16,597	hotpotqa	—	SOB-style grounded extraction
13,303	ChristianAzinn/json-training	sob-ft-json-training	Synthetic schema filler (schema-compliant subset only)
11,300	Team-ACE/ToolACE	sob-ft-toolace	Tool calling
10,630	scrapegraphai_finetuning_subset	—	Grounded JSON extraction
5,209	hermes-function-calling-v1/glaive_func_calling	—	Tool calling
2,052	GTSAIInfraLabSOTAS/DeepJSONEval	sob-ft-deepjsoneval	Hard nested JSON
1,893	hermes-function-calling-v1/func_calling	—	Tool calling
1,340	hermes-function-calling-v1/json_mode_agentic	—	JSON-mode extraction
1,237	hermes-function-calling-v1/json_mode_singleturn	—	JSON-mode extraction
484	paraloq	paraloq-json-extraction-eval	Medical / structured eval → FT rows
472	suneeldk/text-json	sob-ft-text-json	Strict schema-pass subset only

Quality filters applied (vs. raw merge)

Change	Rows removed / kept
Drop all suneeldk/text-json then re-add strict-pass only	−2,000 + 472
Drop hermes-function-calling-v1/func_calling_singleturn	−1,893
Drop ChristianAzinn/json-training schema fail/error rows	−6,677 → 13,303 kept
Scrapegraph combined MinHash dedup (@ 0.85, first-seen-wins)	−18,806 from both scrapegraph sources
Add paraloq eval corpus	+484
Drop remaining extraction schema fail rows (Draft-7 audit)	−54 (48 DeepJSONEval, 6 Hermes json_mode); −45 STAGE-eval after merge
Merge boradorish/STAGE-eval via staged corpus	+19,302 → 19,257 after schema-fail drop

Non-tool extraction rows are 100% schema-compliant under the section (A) validator in audit_sob_ft_extract.py.

Schema

Columns that are entirely null across the merged corpus are dropped. Currently 15 columns:

Column	Description
context	Source text / prior turns (null on ungrounded rows)
question	Extraction or user prompt
json_schema	Target JSON Schema or tool parameter schema
ground_truth	Verified JSON output or tool-call target
validated_output	Validated model output (legacy scrapegraph/hotpotqa rows)
source_dataset	Upstream dataset identifier — use this to filter subsets
record_id	Stable row id within this corpus
question_type	Category / split label when available
question_difficulty	Difficulty or token count when available
schema_complexity	easy / medium / hard (SOB paper rule)
source_id	Upstream id when available
source_answer	Original source answer (legacy rows)
pipeline_version	Pipeline version (legacy rows)
llm_model	Generator model (scrapegraph rows)
source_response_is_valid_flag	Source validation flag (scrapegraph rows)

Removed: candidate_response (always null).

Maintenance scripts

# Audit quality (schema compliance, tool-call parse, dedup stats)
python audit_sob_ft_extract.py --skip-scrapegraph-dedup

# Filter json-training / drop low-quality sources
python filter_sob_ft_extract.py --dry-run

# Drop extraction schema fail/error rows
python drop_schema_fail_rows.py --dry-run

# Scrapegraph combined dedup + push
python dedup_scrapegraph.py --push
python dedup_scrapegraph.py --from-parquet sob-ft-extract-deduped.parquet --push

# Merge paraloq + strict-pass text-json
python merge_paraloq_to_extract.py

# Merge STAGE-eval staged corpus
python merge_ft_corpus.py --corpus mariem123kfg/sob-ft-stage-eval --source-prefix boradorish/STAGE-eval

# Refresh all-null column drop + upload this README
python refresh_sob_ft_extract.py --readme-only

Staged upstream mappers

Script	Source
merge_text_json.py	suneeldk/text-json
merge_deepjsoneval.py	GTSAIInfraLabSOTAS/DeepJSONEval
merge_json_training.py	ChristianAzinn/json-training
merge_toolace.py	Team-ACE/ToolACE
merge_stage_eval.py	boradorish/STAGE-eval
merge_paraloq_to_extract.py	paraloq + text-json strict-pass
merge_hermes_json_mode (1).py	NousResearch/hermes-function-calling-v1
drop_schema_fail_rows.py	Remove Draft-7 fail/error extraction rows

Dataset cards live in dataset_cards/ in the project repo.