feat(sensenova): add validated OCR LoRA training recipe

[leviking98z-rgb]

Summary

• add a validated 32-GPU trainside FlowGRPO recipe for sensenova/SenseNova-U1.5-8B-MoT-SFT using LoRA rank 64 / alpha 128, learning rate 3e-4, six samples per prompt, 320×320 generation, six inference steps, and 300 rollouts
• make the local OCR reward consistently use PaddleOCR plus normalized Levenshtein similarity, with Unicode-aware normalization and edit/substring/exact/counter-IoU diagnostic components
• surface reward-component means during diffusion evaluation and add the short-text train/eval prompt splits used for reward-curve validation
• align the eval dependencies and installation docs with the OCR implementation

Related Issue

N/A

Test Plan

• python3.12 -m compileall -q unirl
• SKIP=no-commit-to-branch pre-commit run --files INSTALL.md datasets/ocr/README.md datasets/ocr/short_text_test.txt datasets/ocr/short_text_train.txt examples/diffusion/sensenova_u1_5/sensenova_u1_5_ocr_trainside_lora_4x8.yaml pyproject.toml requirements.txt setup.py unirl/reward/local/ocr.py unirl/trainer/diffusion.py --show-diff-on-failure
• H20 environment dependency check: PaddlePaddle 3.2.2, PaddleOCR 3.5.0, and python-Levenshtein 0.27.4
• Hydra resolution on an H20 node: python -m unirl.train_diffusion --config-name=diffusion/sensenova_u1_5/sensenova_u1_5_ocr_trainside_lora_4x8 --cfg job --resolve
• one-off assertions (not committed) for LoRA/config values, deterministic eval sampling, Unicode target extraction, normalized Levenshtein behavior, and short-split provenance/filtering
• end-to-end training validation: 4×8 NVIDIA H20 GPUs; sensenova/SenseNova-U1.5-8B-MoT-SFT; the included short-text split; checkpoint checkpoint-134 (optimizer_step=268). On the fixed 32-prompt eval set, mean reward increased from 0.3786 at step 0 to 0.9070 at step 130, and exact match increased from 0.1562 to 0.6562. W&B run: https://wandb.ai/leviking98z-zhejiang-university/tmp/runs/1w1gbj0l

Compatibility / Risk

• The eval extra now installs the dependencies actually used by the local OCR scorer (paddlepaddle, paddleocr, and python-Levenshtein) instead of EasyOCR.
• OCR execution is CPU-backed and may be a throughput bottleneck outside the validated 4×8 setup.
• The new reference recipe requires 32 GPUs by default; users can override paths and logging destinations through the documented environment interpolations.

Reviewer Notes

• AI assistance was used to implement and validate this change; the full diff and generated artifacts were reviewed before publication.
• Duplicate-work check on August 31, 2026 found no open OCR/SenseNova recipe PR or issue. Open PR feat(sensenova): add vLLM-Omni rollout #393 concerns the separate vLLM-Omni rollout backend.
• Checkpoints, W&B files, launcher scripts, and reward-ablation artifacts are intentionally not included.

Checklist

• I reviewed the changed code and removed unrelated/generated artifacts.
• I updated tests, docs, and configs where needed, or explained why not.