Skip to main content
Both of your scripts receive a single PrismContext (ctx). It supplies the metadata and limits your model and loop need, and it controls everything the miner must not: the dataset, the seed, the scoring, and the held-out evaluation. Source: docs/submissions.md:56-74; src/prism_challenge/evaluator/interface.py.

​
PrismContext fields and methods

Source: docs/submissions.md:60-71.

​
What you control - and what you do not

You provide model code and a training loop, not your own data. PRISM supplies and controls the dataset. The miner does not control:
  • the dataset content or splits;
  • the seed and initialization (forced by the harness);
  • the scoring;
  • the held-out evaluation.
Source: docs/submissions.md:72-73; docs/miner/README.md:72-74.

​
Reading the locked data

Read raw text from ctx.data_dir and tokenize it with your own tokenizer or a pre-staged reference. The val/test splits are secret and never exposed to your script - only the challenge scorer reads them. The eval container runs with network=none, HF_HUB_OFFLINE=1, and HF_DATASETS_OFFLINE=1, so there is no network during training. Do not try to download data, tokenizers, or weights at runtime. Fail closed if the locked data is missing rather than fabricating data. Source: docs/miner/README.md:76-81; docs/submissions.md:91-94.

​
Reference tokenizers

ctx.reference_tokenizer(name) loads a pre-staged tokenizer entirely offline. Two references are available: "gpt2" (via a tiktoken cache) and "llama" (via a sentencepiece .model). Using a reference tokenizer never touches the network.
Source: src/prism_challenge/evaluator/interface.py:56-66; docs/miner/README.md:70.
Because the score normalizes by raw UTF-8 bytes, the metric is tokenizer-agnostic - you can bring any tokenizer and still be compared like for like. See Scoring.

​
Distributed geometry

ctx.world_size, ctx.rank, ctx.local_rank, and ctx.device describe the launch shape. The harness launches torchrun --standalone --nnodes=1 --nproc-per-node=<gpu_count> and exposes WORLD_SIZE, RANK, and LOCAL_RANK. Your loop must also work correctly at world_size=1, because the official scored run uses one physical GPU. Source: docs/submissions.md:96-110; docs/scaling.md:22-35. See Constraints for the sandbox, the parameter cap, and the multi-GPU bounds.