lihaoxin2020/evidence-subagent-sft-gpt54-single-all-jina-v1

GPT-5.4 Evidence Subagent SFT with Jina-refreshed Browse Outputs

This dataset contains synthetic SFT conversations for training a small evidence execution subagent for deep-research systems.

Each row is a single delegated evidence-gathering subtask derived from a full DR-Tulu trajectory. GPT-5.4 synthesized the delegated subtask, grouped original tool events into one or more batch tool-call turns, and wrote a structured cited evidence report. Tool outputs are reconstructed from the original trajectory; browse_webpage results were replaced where possible with previously refreshed raw Jina webpage outputs while preserving the original source IDs used by the evidence reports.

Files

• —train.jsonl: SFT rows in conversation format.
• —summary.json: packaging and browse-refresh statistics.
• —trajectory_length_stats.json: full trajectory-length statistics.
• —trajectory_length_stats.md: readable trajectory-length summary.
• —preview_5_replaced_examples.md: five full examples with replaced Jina browse outputs.

Dataset Summary

• —Rows: 10,484
• —Browse tool results seen: 2,968
• —Browse results replaced with refreshed Jina outputs: 2,941
• —Browse results left original due to missing/failed refresh cache: 27
• —Citation-invalid rows after rewrite: 0

Length Summary

Qwen3.5-4B chat-template token counts were computed by rendering each conversation with tokenizer.apply_chat_template(..., tokenize=False, add_generation_prompt=False) and then tokenizing the rendered string.

• —Mean Qwen chat-template tokens: 9,468
• —Median Qwen chat-template tokens: 7,278
• —P90: 14,801
• —P95: 20,591
• —P99: 59,183
• —Max: 304,764
• —Rows over 128K tokens: 13
• —Rows over 262,144 tokens: 2

Structural trajectory lengths:

• —Mean messages including system: 5.63
• —Mean tool rounds: 1.32
• —Mean tool calls: 2.93
• —Median tool calls: 3

See trajectory_length_stats.md and trajectory_length_stats.json for the full distributions.