Convert an autoregressive language model to text diffusion

Do not train, fine-tune, calibrate or select on the validation or test split of any public dataset or benchmark; training splits and other public corpora are fine. That ban also covers eval-split prompts, rewrites of them and teacher-generated answers to them. The examples in /task/public/benchmarks/ are for evaluation only.

Server interface

OpenAI-compatible POST /v1/chat/completions with model=submission, supporting a 40,960-token context (input plus output). Only message.content is scored: return the model's reasoning in reasoning_content (for example with a reasoning parser) or before the final answer in content.

Evaluation

The score is the mean of accuracy/pass@1 across four benchmarks, including GPQA-Diamond and LCB v6, each normalized from random guessing (0) to the original model's performance (1). The evaluation runs 4 identical copies of your server, each on its own H100, and sends requests to them round-robin. Requests use max_tokens=32768 and temperature=1.0 and arrive up to 32 at a time at each copy; the speed gate below is measured on one copy. A reply cut off at max_tokens scores zero. A request with no reply within 3,600 seconds of being sent, including any time it waits behind other requests on its copy, counts as failed and scores zero for its item; failed requests do not stop the evaluation. Scoring stops 22 hours after its first request: requests not answered by then, including any not yet sent, score zero for their items. The benchmarks are sent in the order GPQA-Diamond, MATH-500, MMLU-Pro, LCB v6. For the speed gate, nine of the scored prompts, drawn at random for each evaluation, are sent one at a time with max_tokens=1024 and temperature=1.0; the first is an untimed warm-up. Each of the other eight replies must exceed 450 tokens/second, counting reasoning and answer tokens with the MiniCPM5-2B tokenizer over end-to-end request time.

Artifacts

Take architecture.json's base_model and base_revision from /task/public/STARTING_ARTIFACT.json. In your training evidence, record the datasets, splits and data-generation recipes you used.

Include architecture.json with base_model, base_revision, training_evidence_files and inference_evidence_files. The file lists use paths relative to output/ and include training source, serve.sh and model-loading code. Keep initialization/training logs with the submission as readable text or JSON. Name the source of every training, calibration and selection corpus (in architecture.json or a data manifest) and keep its download commands and logs with the submission. I will check for training-data provenance. Ship compiled dependencies as unmodified PyPI or PyTorch wheels, installed into /tmp at startup. Prewarmed JIT kernel caches (Triton, TorchInductor) may be included.

Submission

Put serve.sh, code, weights and dependencies in /workspace/output/ as regular files. Relative symlinks that resolve inside output/ are copied as the regular files they point to. However, absolute, escaping, dangling or cyclic symlinks fail the submission (for example, use python -m venv --copies, or vendor wheels). The /workspace/output folder is copied to /submission when being evaluated. serve.sh runs as an unprivileged user that can write to /submission, /tmp and the preinstalled Python environment (/opt/venv) and nowhere else; install vendored wheels there or under /submission. You should package dependencies for offline installation, since the evaluation server will not have internet access. The command bash /submission/serve.sh receives HOST=0.0.0.0, PORT=8000 and MODEL_NAME=submission. GET /health must return 200 when ready, within 300 seconds. Startup and inference run offline on one H100 SXM 80GB, with 8 CPU cores and 80 GiB RAM. The serving image is Ubuntu 24.04 with the CUDA 12.8 runtime libraries (no CUDA compiler), Python 3.12 in /opt/venv first on PATH (pip, numpy, pillow, safetensors, httpx preinstalled), gcc and Python headers. PyTorch is not preinstalled: vendor cp312 manylinux wheels (and their dependencies). Every request must answer HTTP 200 with the documented JSON. A request that fails (non-200, connection error, invalid JSON) is retried twice on the same payload. If it still fails, or answers in an undocumented shape, you get no credit for that item. However many requests fail, the evaluation goes on.

Development

The development machine is CPU-only. Use spin --help to rent compute and access model APIs, and spin budget to check how much allowance you have left. The API key your coding harness was started with is for the harness alone. Do not use the API key in your own code, scripts, curl or Spin machines, copy it anywhere, or start extra harness sessions as workers; AI model calls you make yourself must go through spin and spend your allowance. I will check your session transcript for this.

To check your solution locally, run python /task/validate.py --server-url URL. See /task/public for examples and supporting files.

You can check your solution on the same infrastructure I will evaluate it on using frontier selfcheck, which copies /workspace/output/, starts it as the evaluation does, sends public examples through the same response checks and measures the speed requirements on public inputs. Its report gives the stage reached, the digest of the files it checked, the GPU, the startup time and each requirement's measured value and margin, and the end of the server log. You can check your solution four times.

A check usually takes 10-40 minutes (frontier selfcheck status shows this task's limit) and runs while you keep working. See frontier selfcheck --help for more details.

When your solution is final, hand it in with frontier submit: it shows what /workspace/output/ holds and asks you to confirm with frontier submit --confirm, which submits the folder as it is and ends your session. Without a submission, /workspace/output/ is graded as it is when your session ends. See frontier submit --help.

Your session ends the moment you finish a turn without a tool call: background commands, scheduled wake-ups and reminders will not bring you back, and anything not yet in /workspace/output/ is lost. Wait for long jobs with foreground commands under the tool time limit (for example sleep 600; tail -n 3 train.log, repeated) and stop only when the final files are in place.