explcre's picture
4th box eval artifacts: DaTikZ-984 (kimi/vlm x stock/DeTikZify-V2) + MathVision
02147dd verified
|
Raw
History Blame Contribute Delete
2.93 kB

4th box (8xB200) evaluation artifacts

Every per-item file behind the tables in rl/RUN_LOG_4TH_BOX.md of explcre/dsv-experiments (branch gemma-rl-b200-4th). Published so the numbers are checkable and because the source lived on an ephemeral local SSD.

datikz984/ — DaTikZ official 984-item test set

gen_<label>.jsonl — one row per prompt: file_id, caption, gen_code, gen_raw, gold_code. scores_<label>.jsonl — one row per prompt: dsv_mean, dsv_scalar, imagesim_rsim, dreamsim, clipscore, ted_sim, render_graph, gen_compiles, and imagesim_encoder (which encoder scored it).

label model description family ImageSim encoder
D650_kimi_stock / _dtz scale_3brl_blend_rsim_w0p3_detikzify_siglip_termwise @650 (term-wise credit) kimidesc stock / DeTikZify-V2
D650_vlm_stock / _dtz same vlmdesc stock / DeTikZify-V2
CTL650_* scale_3brl_blend_rsim_w0p3_detikzify_siglip @650 (matched control, no term-wise credit) " "
4th_armA (+_dtzenc) dapo_4th_fixabstain_scalar @400 kimidesc stock / DeTikZify-V2
4th_armB (+_dtzenc) dapo_4th_fixabstain_witness_gated @400 kimidesc stock / DeTikZify-V2

R_Sim is not comparable across encoders. On a fixed model, stock -> DeTikZify-V2 moves it by -0.069 (arm A: 0.3892 -> 0.3202) -- roughly 50x the arm-vs-arm differences being measured. Always read the imagesim_encoder field.

Step 650 is used for the D-vs-control pair because it is the newest checkpoint both runs had; comparing across different amounts of training would not be a method comparison.

mathvision/ — MathVision diagram-generation benchmark

eval_results.json — overall + per-category DISTS / CLIP-sim / Edge IoU / Edge F1 / CMMD, and per_image. generation_log.json — per-prompt emitted code and stats.

⚠️ mv_CTL650/generation_log.json holds 730 of 2,920 entries: generation was sharded over 4 GPUs and generate_local.py writes one log per run, so the shards overwrote each other. This affects only code_language_distribution (hence 2,176 unknown). Every headline metric is computed from the PNGs and all are present; the 730 surviving entries are 100% tikz.

renders.tar.gz -- the model's rendered output, one PNG per prompt (2,893 for D, 2,903 for the control). This is what to look at to judge the samples by eye. The DaTikZ side has no equivalent: rendering there happens in-memory during scoring, so its samples are the gen_code field of the generation JSONLs.

Result, in one line

Term-wise credit does not beat its matched control on any benchmark. DaTikZ dsv_mean +0.0007 (kimi, p=0.875) / +0.0016 (vlm, p=0.626); MathVision all four metrics p>=0.16. The only Holm-surviving effect across 10 tests is kimidesc R_Sim at -0.0050 against term-wise credit (Holm p=0.010), unreplicated on vlmdesc.