Qwen Pixel-Edit · SFT & RL dashboard

Instruction-based image editing with executable pixel programs. Teacher-distilled, execution-verified SFT data, then RL. Auto-refresh every 30s.

5,044 / 5,000
accepted trajectories
10,088 / 10,000
estimated SFT rows
10,489teacher attempts
272.2accepted / hour
—ETA to 5k
6,227 / 8,807samples touched
70.7%dataset coverage
2026-11-20 12:52:57 UTClast render

Experiment board

rundatasteplossprogressstatus
v2 · MagicBrushMagicBrush 1,195 0 / 2000.0000
not started
v3 · MagicBrushMagicBrush 3,306 0 / 4000.0000
not started
v3-cont · MagicBrushMagicBrush 3,306 1200 / 12000.4502
done
v4 · MagicBrushMagicBrush 5,082 940 / 16000.4380
stopped
mixed v2 · high-ratio4,109 (ext 21%) 0 / 8000.0000
not started
v5 · mixed-all5,962 (ext 15%) 769 / 20000.5239
stopped
v6 · low-ratio5,522 (ext 8%) 23 / 20000.5308
stopped
rft · best-of-NRFT 5,250 53 / 16000.5532
running
v9 · RFT + target-informedRFT + TI ~5.8k 0 / 16000.0000
not started
v10 · all-data mixed7,619 (MB+UE+Sci) 0 / 20000.0000
not started
v11 · all-data + TI8,793 (MB+UE+Sci+TI) 0 / 20000.0000
not started
v12 · 10k all-data9,447 (RFT+UE+Sci+TI) 56 / 24000.5898
running
v13 · all-data + TI v69,890 (RFT+UE+Sci+TI1.3k) 1042 / 24000.3962
running
v13-warm · from best ckpt9,890 (warm start) 954 / 15000.3905
running
v13-r32 · capacity9,890 (LoRA r32) 933 / 24000.4402
running
v15 · TI 2x dose11,209 (TI 23.5%) 651 / 24000.5007
running
ultraedit-onlyUltraEdit 700 0 / 8000.0000
not started
modelbenchmarktasksexecutedjudge passpass rate
zero-shotMagicBrush dev32 partial 32 26 39.4%
zero-shotUltraEdit dev partial 82 56 56.1%
zero-shotSciForma dev 55 52 2138.2%
v3-cont-1200MagicBrush dev32 32 21 721.9%
v3-cont-1200UltraEdit dev partial 36 21 616.7%
v3-cont-1200SciForma dev 55 43 1934.5%
v4-400MagicBrush dev32 8 7 225.0%
v4-400_part0MagicBrush dev32 8 7 225.0%
teacher deepseekMagicBrush dev32 32 31 1340.6%
teacher deepseekUltraEdit dev 94 87 4851.1%
teacher deepseekSciForma dev 55 54 3360.0%
teacher gpt55MagicBrush dev32 32 28 825.0%
teacher gpt55UltraEdit dev 94 84 4547.9%
teacher gpt55SciForma dev 55 55 4378.2%
teacher gpt56solMagicBrush dev32 32 30 1856.2%
teacher gpt56solUltraEdit dev 94 91 7478.7%
teacher gpt56solSciForma dev 55 50 4785.5%
teacher gpt61MagicBrush dev32 32 29 2165.6%
teacher gpt61UltraEdit dev 94 90 7781.9%
teacher gpt61SciForma dev 55 53 5192.7%
teacher lunaMagicBrush dev32 32 26 1237.5%
teacher lunaUltraEdit dev 94 81 4042.6%
teacher lunaSciForma dev 55 14 1120.0%
teacher seed21proMagicBrush dev32 32 27 1134.4%
teacher seed21proUltraEdit dev 94 72 3739.4%
teacher seed21proSciForma dev 55 51 2952.7%
teacher seed21turboMagicBrush dev32 32 29 825.0%
teacher seed21turboUltraEdit dev 94 73 1718.1%
teacher seed21turboSciForma dev 55 53 1629.1%
teacher terraUltraEdit dev 94 27 1516.0%
teacher terraSciForma dev 55 43 3869.1%

Data quality · RFT & target-informed distillation

experimentnjudge passCLIP gain vs targetL1 gain vs targetnotes
source-only accepted (base)5,826—-0.0013—all accepted trajectories
RFT best-of-N selected5,826—-0.0000—1,691 multi-candidate; 3,012 positive
STaR target-informed (gpt55)10034+0.0035-0.0011flip-reuse 73/100; 0 hardcoded (refined filter)
source-only matched base100—+0.0011-0.0039accepted trajectories only

Official benchmark metrics (vs target)

modelbenchmarknL1 ↓PSNR ↑SSIM ↑LPIPS ↓CLIP-I ↑CLIP gain ↑DINO ↑DINO gain ↑
sft_magicbrush_27b_oneshot_v3_cont-1200MagicBrush dev32 210.078718.90 0.8180.213 0.906-0.02340.000+0.0000
v3cont1200 sciformaSciForma dev 430.021226.01 0.9540.066 0.984-0.00690.000+0.0000
zero sciformaSciForma dev 520.010825.13 0.9590.067 0.987-0.00300.000+0.0000
teacher gpt55MagicBrush dev32 280.061918.57 0.8330.190 0.928-0.00270.000+0.0000
teacher gpt61MagicBrush dev32 290.062819.09 0.8240.192 0.921-0.00600.000+0.0000

SFT · Qwen3.8-27B one-shot

0 / 400training step
0.0000last loss
0.0 GBpeak memory
1,195train rows
3/32 = 9.4%zero-shot baseline (dev32)
checkpointdev tasksparsedexecutedjudge passpass rate
waiting for first checkpoint

SFT · Qwen3.8-27B one-shot (code-only)

0 / 400training step
0.0000last loss
0.0 GBpeak memory
1,719train rows
3/32 = 9.4%zero-shot baseline (dev32)
checkpointdev tasksparsedexecutedjudge passpass rate
waiting for checkpoint-200

Per-teacher contribution

teacherattemptsexec okjudge passacceptedaccept rateexec failteacher error
gpt616186 598745014537 73.3%13861
terra1685 1572335337 20.0%1112
gpt551268 11938385 6.7%750
luna1350 11528585 6.3%17424

Recent accepted

sampleinstructionteacherscores
magicbrush_train_398489_2 put a cute puppy beside the catgpt61 5/5/4
magicbrush_train_10936_3 let there be bouldersgpt61 4/5/3
magicbrush_train_303651_3 let the girl wear formal clothesgpt61 4/4/3
magicbrush_train_566188_2 let the woman hold a blow dryergpt61 4/4/3
magicbrush_train_188109_3 remove the small chairluna 4/4/3
magicbrush_train_249131_3 Can it be fries next to the cup?gpt55 4/4/3
magicbrush_train_398489_1 make the cat happygpt61 4/5/4
magicbrush_train_159282_1 Add an ashtray next to the vasegpt61 5/5/4
magicbrush_train_303651_1 let the dog bite the girlgpt61 4/4/3
magicbrush_train_548913_1 Let the person load the cake with a spoon.gpt61 4/5/3

Dataset coverage

Touched 6,227 of 8,807 MagicBrush train rows across 51 shards (70.7%). Accepted 5,044 trajectories; target 5,000 trajectories / 10,000 SFT rows.