zzqsb/f2llm-tevatron-jsonl

F2LLM — Tevatron JSONL reformat

A reformat of the public codefuse-ai/F2LLM embedding-training dataset into the Tevatron "legacy" dense-retriever training JSONL format. No new data is introduced — only the schema is converted, plus a retrieval-only subset is sampled.

Files

file	rows	description
f2llm_train.jsonl	5,933,964	Full conversion of all F2LLM task types (retrieval + classification + clustering).
f2llm_680k_train.jsonl	680,000	Retrieval-only random subset (seed 42), reservoir-sampled from the retrieval rows; sized to match RLHN-680k for retriever training.
f2llm_lenbalanced_680k_train.jsonl	680,001	Retrieval-only, query-length-balanced subset (seed 42): equal mass (226,667 each) in ModernBERT-token bins short <100 / medium [100,300) / long [300,∞). Same total size as f2llm_680k so only the length distribution changes (natural ≈79/15/6% → 33/33/33%).
f2llm_lenbal6633_680k_train.jsonl	680,001	Retrieval-only, two-bin 66/33 subset (seed 42): split at 100 ModernBERT tokens, <100 = 66.7% (453,334) / >=100 = 33.3% (226,667); a softer rebalance between the natural short-skew and the fully-balanced version.

Schema (one JSON object per line)

{
  "query_id": "0",
  "query": "Instruct: ...\nQuery: ...",
  "positive_passages": [{"docid": "<md5>", "title": "", "text": "..."}],
  "negative_passages": [{"docid": "<md5>", "title": "", "text": "..."}]
}

• docid is an md5 hash of the passage text; title is empty.
• Up to 24 F2LLM negatives are down-sampled to neg_k=7 per row; retrieval rows carry 7 negatives.
• Query text keeps F2LLM's inline Instruct:/Query: instruction prefix.

How it was built

Converted with the prepare_f2llm.py convert_row logic (F2LLM parquet → Tevatron JSONL), streamed with bounded pyarrow.iter_batches. The retrieval-only subset filters rows by F2LLM's infer_task_type (query == retrieval) and reservoir-samples 680,000 rows with seed 42.

The two length-controlled subsets (f2llm_lenbalanced_680k, f2llm_lenbal6633_680k) were built with make_f2llm_lenbalanced_subset.py: a single streaming pass that filters retrieval rows, tokenizes each query with the ModernBERT tokenizer (cap 512), bins by token length, and reservoir-samples per bin (seed 42). They exist to test whether a retrieval model's short-query behavior is driven by the training length distribution rather than the loss.

Attribution & license

Derived from codefuse-ai/F2LLM (public). Please refer to the original dataset for source licensing and terms; this repository only reformats that public data.