zen-E's picture
Update PLatour usage for the unified runtime
75ff6d2 verified
|
Raw History Blame Contribute Delete
2.45 kB
metadata
pretty_name: PLatour-DeepCoder-Data
tags:
  - platour
configs:
  - config_name: deepcoder
    data_files:
      - split: train
        path: deepcoder/train.jsonl
      - split: validation
        path: deepcoder/validation.jsonl
  - config_name: lcb
    data_files:
      - split: test
        path: deepcoder/lcb.jsonl
  - config_name: humaneval
    data_files:
      - split: test
        path: deepcoder/humaneval.jsonl
  - config_name: mbpp
    data_files:
      - split: test
        path: deepcoder/mbpp.jsonl
  - config_name: humaneval_template_adapted
    data_files:
      - split: test
        path: deepcoder/humaneval_template_adapted.jsonl
  - config_name: mbpp_template_adapted
    data_files:
      - split: test
        path: deepcoder/mbpp_template_adapted.jsonl

PLatour-DeepCoder-Data

Frozen PLatour task inputs. File bytes and question order are preserved for exact SHA-256 checks.

TRAIN/VAL checker cases are rebuilt from the pinned public DeepCoder revision using python -m qwen35.data --output <fresh-case-directory>. The original case archives are not included. LCB v6 checker cases are obtained separately with qwen35.judge_cases.

Template-adapted HumanEval/MBPP prompts have separate subsets and must be labeled as such in results.

Use with PLatour

For the project training/evaluation CLI, follow the installation guide and activate the project environment. From the checkout root, download and verify one experiment's inputs:

platour-exps download-data --experiment 26

Training and evaluation also download missing signed inputs automatically. Generation uses the active Python; no separate decoder path is required. See the task recipes and data guide for all experiments, split contracts and offline use.

Programming evaluation also requires the coding judge setup.

Sources and use

Consult the manifest for source identities and checksums and the project task documentation for prompt and scoring contracts. This collection combines upstream datasets with different licenses. Their original terms continue to apply; the project code license does not replace the source dataset licenses.