Codeseys/composer-replication-framework

1# CITATION.cff — Citation File Format2# https://citation-file-format.github.io/3# Used by HF, GitHub, Zenodo to render a "Cite this repository" UI.4 5cff-version: 1.2.06message: "If you use this framework or its derivative artifacts, please cite as below."7type: software8title: "Composer 2.5 Replication Framework: Methodology and Integration Architecture for Open Replication of Cursor's Agentic Coding Recipe"9abstract: >10  An open-source methodology and integration architecture for replicating11  Cursor's Composer 2.5 recipe on a HuggingFace base model, plus a novel12  multi-teacher trace-replay distillation reward channel that complements13  the published SDPO/OPSD method (which Cursor's "Targeted RL with Textual14  Feedback" uses). Pre-experimental v0.0 release: methodology paper, audited15  recipe mapping, integration architecture across TRL/VeRL/OpenEnv,16  empirical economic-feasibility result for the novel channel ($0.98/trace),17  and a working code skeleton with 38 passing unit tests.18 19authors:20  - family-names: "Codeseys"21    given-names: ""22    affiliation: "Independent researcher"23    # Replace with real ORCID if available:24    # orcid: "https://orcid.org/0000-0000-0000-0000"25 26repository-code: "https://huggingface.co/Codeseys/composer-replication-framework"27url: "https://huggingface.co/Codeseys/composer-replication-framework"28date-released: "2026-05-25"29version: "0.0.0"30license: "MIT"31 32keywords:33  - reinforcement-learning34  - post-training35  - distillation36  - agentic-coding37  - composer-2.538  - cursor39  - kimi-k240  - grpo41  - dapo42  - sdpo43  - opsd44  - trl45  - verl46  - openenv47  - llm48 49# Primary upstream works this framework depends on / cites50references:51  - type: article52    title: "Introducing Composer 2.5"53    authors:54      - name: "Cursor Team"55    year: 202656    url: "https://cursor.com/blog/composer-2-5"57 58  - type: article59    title: "Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models"60    authors:61      - family-names: "Zhao"62        given-names: "Siyan"63      - family-names: "Xie"64        given-names: "Zhihui"65      - family-names: "Liu"66        given-names: "Mengchen"67      - family-names: "Huang"68        given-names: "Jing"69      - family-names: "Pang"70        given-names: "Guan"71      - family-names: "Chen"72        given-names: "Feiyu"73      - family-names: "Grover"74        given-names: "Aditya"75    year: 202676    url: "https://arxiv.org/abs/2601.18734"77    notes: "OPSD — single-LLM self-distillation; provides the reference loss implementation lifted by this framework."78 79  - type: article80    title: "Reinforcement Learning via Self-Distillation"81    authors:82      - family-names: "Hübotter"83        given-names: "Jonas"84      - family-names: "Lübeck"85        given-names: "Frederike"86      - family-names: "Behric"87        given-names: "Lejs"88      - family-names: "Baumann"89        given-names: "Anton"90      - family-names: "Bagatella"91        given-names: "Marco"92      - family-names: "Marta"93        given-names: "Daniel"94      - family-names: "Hakimi"95        given-names: "Ido"96      - family-names: "Shenfeld"97        given-names: "Idan"98      - family-names: "Buening"99        given-names: "Thomas Kleine"100      - family-names: "Guestrin"101        given-names: "Carlos"102      - family-names: "Krause"103        given-names: "Andreas"104    year: 2026105    url: "https://arxiv.org/abs/2601.20802"106    notes: "SDPO — formalizes the same mechanism as Cursor's Targeted RL with Textual Feedback. ICLR 2026 Scaling Post-training Workshop."107