parser_ratio — measure the SG-parser effect (gpt-4o-mini vs GPT-5.4) on GED

Purpose: all published CEDI rows had their transcripts parsed into unique_sg with gpt-4o-mini. The Gemini-2.5-Pro SVG-500 run was parsed with a different model, so its GED / GED_hal / GED_cov are not directly comparable. Re-parsing this fixed sample with GPT-5.4 lets us measure the parser-induced ratio and correct for it.

Input

sample60_parsed_gpt4omini.json — 60 SVG-500 images (24 COCO / 12 VG / 24 ADE20K), drawn from the Qwen2.5-VL-7B-Instruct v19 run (GPT-4o examiner backbone). Each entry has:

image_id
sg — ground-truth scene graph (do not modify)
conversations — merged transcript, ~25 turns/image, each with question / response
unique_sg_gpt4omini — the existing gpt-4o-mini parse (reference; do not modify)
dist_score_gpt4omini — GED under that parse (mean over the 60 = 122.90)

What to run

Re-parse the same conversations with GPT-5.4 using the identical parser prompt and post-processing. Everything except the model must stay the same:

Prompt: PROMPT.txt (verbatim; {question} is substituted with one turn's full response string — the batch parser passes the whole response as a single unit, not sentence-split).
System message: "From the given sentence, your task is to extract meaningful triplets formed as <subject, predicate, object>."
Post-processing: refine_output() in grader/sg/llm_parser.py — drops non-physical words, keeps <s, p, o>, <o, is, attr> and bare <o>; triples are serialized as "( a , b , c )" and de-duplicated into a set.
Temperature / decoding: whatever the standard LLMChat default is; keep it consistent.

Do NOT recompute GED — that step is LLM-free and will be run on our side.

Expected output

sample60_parsed_gpt54.json: same 60 entries, same image_id order, each adding unique_sg_gpt54 (list of "( a , b , c )" strings). Keeping the other fields is fine.

Where to put it

Upload to this HF dataset under parser_ratio/sample60_parsed_gpt54.json.

We will then compute GED / GED_hal / GED_cov under both parses on the same 60 images and report the parser ratio (gpt-4o-mini ÷ GPT-5.4) alongside the already-measured examiner backbone ratio.