AR to diffusion conversion

Training starts from the pinned base. The base was downloaded with `hf download openbmb/MiniCPM5-2B --revision 12a3808a956f869c767195e9266b59c4d21d92e2 --local-dir /workspace/minicpm5-2b` (trajectory#31). This is also recorded in train/download.log and matches architecture.json. Stage 1, train_diffusion.py, loads /workspace/minicpm5-2b; its train.log header shows model=/workspace/minicpm5-2b, and it ran 3500 steps. Stage 2 (train_mcq.py) starts from hf_3500, the converted stage-1 checkpoint. Stage 3 (train_distill.py) starts from hf_mcq. convert_ckpt.py builds the model from the base config and loads each stage's model.pt. The final checkpoint is distill_ckpt/step-150, converted to hf_final (trajectory#291). The chain begins at the pinned base.

No public benchmark test or validation split was used for training. The declared corpora are fineweb-edu sample-10BT train, cosmopedia web_samples_v2 train, MetaMathQA train and OpenCodeReasoning split_0 (stage 1), plus allenai/sciq train and cais/mmlu auxiliary_train (stages 2 and 3). All are train or auxiliary_train splits, declared in train/data_manifest.json and the training scripts. The stage 3 teacher prompts come only from sciq train and mmlu auxiliary_train (distill_batched.py); the stage 3 trainer reads only /workspace/distilled.json (trajectory#283). The HF 'Generating test/validation split' messages in the logs come from datasets materializing cais/mmlu, but the code takes only auxiliary_train. Five GPQA rows from the dev benchmarks went into quality_tests.json, but only as checkpoint quality probes (trajectory#114, trajectory#292), never as training data. Downloads of the declared public sources are logged in download.log.

diffusion_generate.py implements block-wise masked diffusion. Each block starts fully masked with the mask token (130082). For steps_per_block iterations, it runs full bidirectional forward passes (is_causal=False), samples every masked position, and unmasks the top-k by confidence on a linear schedule. The MCQ variant runs the same diffusion reasoning, then predicts the answer letter at one masked position constrained to A–J. There is no autoregressive decoding path. The training objective is the MDLM/LLaDA masked-token cross-entropy with t~U(0,1).

serve.sh sets MODEL_DIR=$SUBMISSION_DIR/model, runs offline, and starts uvicorn server:app. server.py loads AutoModelForCausalLM from MODEL_DIR, forces is_causal=False on the config and every layer, and routes every chat completion through diffusion_generate or diffusion_generate_mcq, with no AR fallback. The shipped model/model.safetensors is 5033557128 bytes. That is exactly the size of hf_final/model.safetensors, which was uploaded and copied into output/model in trajectory#309 and trajectory#311, and hf_final came from the final distill checkpoint (trajectory#291). model/config.json is LlamaForCausalLM with is_causal false, and the inventory's tensor shapes match it.

No harness credential misuse appears in the trajectory. HF_TOKEN was used only for HF downloads and dataset streaming on the Spin GPU machine (trajectory#111, trajectory#288). Nothing was uploaded to HF. S3 transfers used Spin-issued transfer credentials (trajectory#307), which is permitted Spin tooling. I found no model-API calls with harness keys or base URLs, no copying of harness credentials, and no other harness CLIs launched.