mariem123kfg/sob-ft-finetuning

SOB Fine-Tuning Dataset (Extraction + Correction)

Internal fine-tuning corpus for JSON extraction and JSON repair tasks, built to fine-tune our best-performing model from the CAREBench benchmark on both scenarios it's evaluated on.

What this is

100,405 rows split into two paired tasks:

Split	Rows	Task	What the model sees	What it should produce
extraction	50,204	JSON extraction	context, question, json_schema	A valid JSON object matching the schema (validated_output)
correction	50,201	JSON repair	Same inputs, plus a corrupted JSON (errored_json)	The corrected JSON (validated_output), same target as extraction

train (100,405 rows) is the two splits combined, for training a single model on both tasks together.

from datasets import load_dataset

extract = load_dataset("mariem123kfg/sob-ft-finetuning", split="extraction")
correct = load_dataset("mariem123kfg/sob-ft-finetuning", split="correction")
all_rows = load_dataset("mariem123kfg/sob-ft-finetuning", split="train")

How it was built

1. Merge. 12 upstream sources (hotpotqa, scrapegraph web-scrape extraction, STAGE-Eval, ToolACE, Hermes function-calling, DeepJSONEval, paraloq, and others) were mapped onto a shared schema (context/question/json_schema/ground_truth) and combined.

2. Quality filter. Every row was validated with a Draft-7 JSON Schema checker against its own ground_truth. Rows that failed (schema-noncompliant labels, e.g. missing required fields, wrong types) were dropped rather than kept as "correct" training examples, since training on invalid ground truth would teach the model the wrong thing about what compliant output looks like. This removed sources/rows with low compliance rates outright (an entire low-quality source at ~24-29% compliance, a tool-calling variant at ~49%) and trimmed the largest remaining source down to its schema-valid subset only.

3. Deduplication. The two largest sources (both derived from the same underlying scrapegraph web-scrape pipeline) had substantial internal near-duplication, over 20% within each individually. These were deduplicated in a single combined MinHash/LSH pass (similarity threshold 0.85, first-seen-wins) so the corpus doesn't overweight one generation pipeline's idiosyncrasies.

4. Volume additions. Two datasets originally built as held-out evaluation sets (paraloq-json-extraction-eval, STAGE-eval) were deliberately repurposed into training data after review, since they were schema-compliant and added useful volume/diversity. Important: those standalone eval repos should no longer be treated as clean/unseen benchmarks for any model trained on this corpus.

5. Error injection (correction split only). For the correction task, a copy of each clean, schema-valid validated_output was algorithmically corrupted (errored_json) to create repair training pairs. Errors are injected with controlled type, count, and difficulty per row:

• Error types: e.g. string_truncation, number_as_string, case_corruption, type_break, array_duplicate, drop_required_field, sibling_value_swap, numeric_off_by_one, word_typo. Multiple error types can be combined in a single row (error_type records them concatenated, error_types as a list).
• Error count: 0-3 injected errors per row, tracked in error_count.
• Error difficulty: each individual injected error is tagged easy/medium/hard (error_difficulty/error_difficulties), letting training and eval slice by how subtle the corruption is.
• Error location: the exact JSON path(s) affected are recorded (error_location/error_locations), along with a human-readable description of what changed (error_description/error_descriptions) and a structured combined record (errors_json).
• error_source marks all injected errors as algorithmic (rule-based corruption, not model-generated), for reproducibility.

Extraction-split rows carry error_count == 0 by construction (errored_json == validated_output, i.e. no corruption applied), so the same schema works across both splits without special-casing.

Columns

Column	Extraction	Correction
record_id	✓	✓
context, question, json_schema	✓	✓
validated_output	✓ (target)	✓ (repair target)
errored_json	✓ (= target, no corruption)	✓ (corrupted input)
error_count, error_difficulty, error_type(s)	✓ (zero/none)	✓
error_location(s), error_description(s), errors_json	optional	✓
source_dataset, schema_complexity, question_type	✓	✓

Source mix

Rows	source_dataset
35,025	scrapegraphai_full
19,257	boradorish/STAGE-eval
16,597	hotpotqa
13,303	ChristianAzinn/json-training
10,630	scrapegraphai_finetuning_subset
2,052	GTSAIInfraLabSOTAS/DeepJSONEval
1,340	hermes-function-calling-v1/json_mode_agentic
1,237	hermes-function-calling-v1/json_mode_singleturn
484	paraloq
472	suneeldk/text-json
8	Team-ACE/ToolACE

Eval firewall

Holdout	Overlap with this dataset
interfaze-ai/sob test split (~5k rows)	0 shared record_ids
MoetezF/SOB-with-errors-injection test split (~5k rows)	0 shared record_ids

Both held-out sets were checked for record_id overlap before publishing and confirmed clean. Do not merge either eval set into this training corpus, doing so would invalidate any CAREBench numbers reported against them going forward.