ahmedheakl/rand-mobile-grpo

Mobile-O — GRPO checkpoint (v2mcptf-grpo-500)

The project's main GRPO checkpoint: 500 steps of GRPO from the v2mcptf-mixed SFT init, with a unified OCR + GenEval reward. This is the baseline every later experiment was measured against.

For the best overall checkpoint see ahmedheakl/rand-mobile (soup3-targets), which meets 3 of 6 targets to this one's 2. This checkpoint is nonetheless part of that soup's lineage — it is one of the two task vectors inside ta-dpg-s1.0, which is one of the three soup members.

Benchmarks

Best setting is cfg 3.0, 20 DPM-Solver++ steps (2 of 6 targets met):

benchmark	cfg 3.0 (best)	cfg 1.5 (protocol)	target	status @ cfg 3.0
GenEval	0.9284	0.9116	≥ 0.90	met
ImageReward (MJHQ-30K)	0.9250	0.7633	≥ 0.90	met
DPG-Bench	81.727	80.679	≥ 85	not met
FID (MJHQ-30K)	15.398	13.508	≤ 8	not met
ImgEdit (local Qwen2.5-VL-72B judge)	3.020	3.093	≥ 3.5	not met
GEdit (EN, local Qwen2.5-VL-72B judge)	6.620	6.520	≥ 6.7	not met

Other measured settings on this checkpoint: cfg 4.5 gives GenEval 0.9237 / DPG 82.331 / ImageReward 0.9727 (its best) but FID 17.165 and GEdit 6.420. Interval guidance (CFG applied only for t ∈ [0, 0.9]) at cfg 4.5 gives DPG 82.979 — this checkpoint's best DPG — at FID 15.959.

Note the FID/alignment trade: cfg 1.5 has the best FID (13.508) and the worst ImageReward (0.7633); cfg 4.5 inverts it. No setting of this checkpoint clears more than two targets.

What is in this repo

Only the trained head: the SANA DiT and the diffusion connector, 602 tensors (model.dit.* 548, model.diffusion_connector.* 54). The vision-language model is frozen during training and is not included here.

To run this you also need:

• openbmb/MiniCPM-V-4_6 — the frozen VLM encoder
• Efficient-Large-Model/Sana_600M_512px_diffusers — for the DC-AE (f32, 32-channel, 16×16 latent at 512px) and the scheduler config

Connector type is mcptf, fusing 1 VLM layer (derive from fusion.layer_weights.shape[0]; a vlm_num_layers field elsewhere in this project can incorrectly read 4 — the weights are authoritative).

Inference settings used for every number above

scheduler	DPM-Solver++, solver_order=2, flow_shift=3
steps	20
guidance	cfg 3.0 for the best row, cfg 1.5 for the protocol row
null condition	the empty prompt through the same VLM+connector path (not a zero vector)
resolution	512×512

For editing, the null condition is the instruction rather than the empty prompt.

Training

GRPO from SFT init Mobile-O-0.5B-SFT-minicpm-v2mcptf-mixed, 500 steps.

reward	unified OCR + GenEval
group size	12
learning rate	6e-5
KL beta	0.04
clip range	1e-4
advantage clip	5.0
rollout	10 denoise steps, cfg 1.0, shift 3.0, noise level 0.7

Caveats

• Four of six targets are not met (DPG, FID, ImgEdit, GEdit). This is the project baseline, not its best result.
• GEdit and ImgEdit are scored by a local Qwen2.5-VL-72B judge, not the GPT-4o leaderboard scale; the two are not comparable.
• The head alone is not a runnable model — see What is in this repo.