Skip to content

THIRD_PARTY_NOTICES.md

Latest commit

 

History

History
114 lines (82 loc) · 4.44 KB

THIRD_PARTY_NOTICES.md

File metadata and controls

114 lines (82 loc) · 4.44 KB
Raw

Third-Party Notices

This repository (C3) is built on top of, and interoperates with, third-party open-source software and public datasets. This document summarizes major attributions. It is not a substitute for the full license texts of those projects.

Important: This repository does not redistribute third-party datasets. Dataset files are downloaded by scripts under scripts/data/ from their official sources and prepared into canonical JSONL artifacts.

For the single source of truth on dataset IDs, pinned revisions, prepared artifact paths, and SHA256 checksums, see:

  • configs/data_manifest.yaml
  • docs/DATA_SOURCES.md

Upstream base: OpenRLHF

  • Project: OpenRLHF (OpenRLHF/OpenRLHF)
  • License: Apache License 2.0
  • Notes: We preserve openrlhf/ for traceability and compatibility. See docs/UPSTREAM.md.

Major software dependencies (non-exhaustive)

C3 relies on the standard LLM training ecosystem, including (but not limited to):

  • Ray (distributed execution)
  • vLLM (fast inference engine)
  • DeepSpeed (distributed training)
  • PyTorch (core ML framework)
  • Hugging Face Transformers / Datasets / Hub tooling

Each dependency is governed by its own license; please refer to their upstream repositories and license files.

Datasets used by the paper configs (downloaded; no redistribution)

Below items are summarized exactly as pinned in configs/data_manifest.yaml (dataset IDs / splits / outputs). The manifest also includes pinned upstream revisions and prepared-artifact SHA256 values for reproducibility.

MATH (train) + MATH500

  • Academic origin: Hendrycks et al., “Measuring Mathematical Problem Solving With the MATH Dataset”
  • Upstream HF IDs (pinned in manifest):
    • qwedsacf/competition_math (used for the train artifact in this release; see manifest notes)
    • HuggingFaceH4/MATH-500 (benchmark)
  • Prepared outputs (via scripts/data/prepare_math.py):
    • data/MATH/train.jsonl
    • data/MATH500/test.jsonl
  • License: see the upstream dataset cards; C3 does not redistribute raw data.

GSM8K

  • Academic origin: Cobbe et al., “Training Verifiers to Solve Math Word Problems”
  • Upstream HF ID (pinned in manifest): openai/gsm8k (config: main)
  • Prepared output:
    • data/GSM8K/test.jsonl
  • License: see the upstream dataset card; C3 does not redistribute raw data.

CMATH (train/test)

  • Academic origin: CMATH benchmark (see upstream dataset card for citation details)
  • Upstream HF ID (pinned in manifest): weitianwen/cmath
  • Prepared outputs (via scripts/data/prepare_math.py):
    • data/CMATH/train.jsonl (from upstream validation split, per manifest pin/notes)
    • data/CMATH/test.jsonl
  • License: see the upstream dataset card; C3 does not redistribute raw data.

HumanEval

  • Benchmark: HumanEval (see upstream dataset card for citation details)
  • Upstream HF ID (pinned in manifest): openai/openai_humaneval
  • Prepared output (via scripts/data/prepare_code.py):
    • data/HumanEval/test.jsonl
  • License: see the upstream dataset card; C3 does not redistribute raw data.

APPS

  • Benchmark: APPS (see upstream dataset card for citation details)
  • Upstream HF ID (pinned in manifest): codeparrot/apps
  • Prepared output (via scripts/data/prepare_code.py):
    • data/APPS/test.jsonl
  • License: see the upstream dataset card; C3 does not redistribute raw data.

MBPP + MBPP+ (EvalPlus)

  • Academic origin: MBPP benchmark (see upstream dataset card for citation details)

  • Upstream HF ID (pinned in manifest): google-research-datasets/mbpp (config: full)

  • MBPP prepared outputs (via scripts/data/prepare_code.py):

    • data/MBPP/train.jsonl
    • data/MBPP/test.jsonl

  • MBPP+ (stronger tests) is prepared from EvalPlus when pinned and available:

    • Manifest entry name: MBPP+
    • Manifest source kind: evalplus
    • Manifest pinned provenance tag: evalplus@0.3.1
    • Prepared output (via scripts/data/prepare_code.py):
      • data/MBPP_PLUS/test.jsonl

Note on MBPP+ reproducibility:

  • This release pins MBPP+ to an explicit EvalPlus version tag in the manifest.
  • Strict verification (--strict 1) enforces that your environment matches the pinned provenance rule.
  • See docs/DATA_SOURCES.md for strict-mode behavior.

Acknowledgements

If you believe an attribution is missing or incorrect, please open an issue or PR with:

  • the upstream project/dataset,
  • a pointer to its license / dataset card,
  • and where it is used in this repository.