{"id":32460338,"url":"https://github.com/vila-lab/ptts","last_synced_at":"2025-10-26T11:56:08.857Z","repository":{"id":319246366,"uuid":"1074954561","full_name":"VILA-Lab/PTTS","owner":"VILA-Lab","description":"P-TTS: inference-time reasoning data augmentation that scales prompt-space with principled instructions, matching/exceeding larger-data baselines.","archived":false,"fork":false,"pushed_at":"2025-10-17T07:49:58.000Z","size":1975,"stargazers_count":3,"open_issues_count":0,"forks_count":0,"subscribers_count":0,"default_branch":"main","last_synced_at":"2025-10-18T10:43:08.195Z","etag":null,"topics":[],"latest_commit_sha":null,"homepage":"","language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"apache-2.0","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/VILA-Lab.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2025-10-12T19:40:36.000Z","updated_at":"2025-10-17T23:48:22.000Z","dependencies_parsed_at":"2025-10-18T15:33:37.135Z","dependency_job_id":null,"html_url":"https://github.com/VILA-Lab/PTTS","commit_stats":null,"previous_names":["vila-lab/ptts"],"tags_count":null,"template":false,"template_full_name":null,"purl":"pkg:github/VILA-Lab/PTTS","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/VILA-Lab%2FPTTS","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/VILA-Lab%2FPTTS/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/VILA-Lab%2FPTTS/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/VILA-Lab%2FPTTS/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/VILA-Lab","download_url":"https://codeload.github.com/VILA-Lab/PTTS/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/VILA-Lab%2FPTTS/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":281099981,"owners_count":26443537,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2022-07-04T15:15:14.044Z","status":"online","status_checked_at":"2025-10-26T02:00:06.575Z","response_time":61,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":[],"created_at":"2025-10-26T11:56:08.022Z","updated_at":"2025-10-26T11:56:08.848Z","avatar_url":"https://github.com/VILA-Lab.png","language":"Python","funding_links":[],"categories":[],"sub_categories":[],"readme":"# P‑TTS: Prompting Test‑Time Scaling 🚀\n\n**90 examples can beat 1K** — P‑TTS uses principled **instructional prompt augmentation** to turn **90 AIME seeds** into **900 high‑utility training examples**, delivering strong reasoning with far less data.\n\n\u003cp align=\"center\"\u003e\n  •\n  📄 \u003ca href=\"https://arxiv.org/abs/2510.09599\"\u003ePaper\u003c/a\u003e •\n  📊 \u003ca href=\"https://huggingface.co/datasets/P-TTS/P_TTS-Full\"\u003eDataset\u003c/a\u003e\n\u003c/p\u003e\n\n\u003cp align=\"center\"\u003e\n  \u003cimg src=\"image/Results.png\" alt=\"P-TTS overview\" width=\"80%\"\u003e\n\u003c/p\u003e\n\n## Table of Contents\n- [What is P-TTS?](#what-is-p-tts)\n- [Key Results](#key-results)\n- [Training Data](#training-data)\n- [Training](#training)\n- [How It Works (pipeline)](#how-it-works-pipeline)\n- [Reproduce](#reproduce)\n- [Citation](#citation)\n\n## What is P‑TTS?\n\nP‑TTS expands a small, vetted seed set (90 AIME 2022–2024 problems) by **wrapping** each problem with *principled instructions* to elicit diverse reasoning traces from a teacher model (DeepSeek‑R1). We then fine‑tune Qwen2.5‑Instruct models on these augmented traces.\n\n**Principles used (unchanged question text; wrappers are prefixed/suffixed):**\n\n* **Reward** – e.g., \"I'll tip \\$200,000 for a better solution!\"\n* **Penalty** – \"You will be penalized if the answer is wrong.\"\n* **Correctness** – \"You MUST provide the correct answer.\"\n* **Step‑by‑Step** – \"Think step by step.\"\n\n\u003e Data scales via augmentation multipliers m ∈ {1, 4, 5, 10}: **90 → 360 → 450 → 900**.\n\n## Key Results\n\n**Benchmarks:** AIME24, AIME25, MATH500, GPQA‑Diamond.\n**Backbone:** Qwen2.5‑Instruct (7B/14B/32B).\n**Metric:** accuracy (lm‑evaluation‑harness; greedy decoding).\n\n| Model         | #Train ex. |     AIME24 |     AIME25 |    MATH500 |     GPQA‑D |       Avg. |\n| ------------- | ---------: | ---------: | ---------: | ---------: | ---------: | ---------: |\n| **P‑TTS‑32B** |        900 | **73.33%** | **53.33%** | **94.20%** | **60.61%** | **70.35%** |\n| **P‑TTS‑14B** |        900 |     53.33% |     26.67% |     90.40% |     51.01% |     55.35% |\n| **P‑TTS‑7B**  |        900 |     43.33% |     26.67% |     84.20% |     41.92% |     49.03% |\n\n## Training Data\n\nThe training dataset consists of 900 high-quality reasoning examples generated from 90 AIME seed problems. Each seed problem is augmented using principled instruction wrappers and processed through DeepSeek-R1 to create diverse reasoning traces.\n\n**Dataset Composition:**\n- **Source**: 90 AIME problems (2022-2024)\n- **Augmentation**: 4 instruction wrapper types with reward variants\n- **Final Size**: 900 training examples\n\n### Data Tokenization\n\nBefore training, you need to tokenize your raw dataset. Use the provided tokenization script:\n\n```bash\n# Run the tokenization script\npython tokenize_data.py\n```\n\n\n## Training\n\nTo run training, you can find our script at `train/sft.py` which you can invoke via one of the `train/sft*.sh` scripts, or you can launch via `train/launch.sh` if you are on a SLURM cluster (requires editing the file for your cluster setup).\n\n### Configuration\n\n**Hardware Requirements:**\n- **For 7B models**: 4x A100 GPUs\n- **For 32B models**: 6x B200 GPUs\n\n**Quick Start:**\n```bash\ngit clone https://github.com/simplescaling/s1.git\ncd s1\npip3 install -r requirements.txt\n# First tokenize your data\npython tokenize_data.py\n# Then run training\nbash train/sft.sh\n```\n\u003e Note: Training scripts are adapted from [simplescaling/s1](https://github.com/simplescaling/s1) (Apache-2.0).\n\n### Training Data\n\nThe script expects your training data in CSV format. Update the `train_file_path` variable in `sft.sh`:\n```bash\n--train_file_path=\"xx_tokonized.csv\"\n```\n---\n\n## How It Works (pipeline)\n\n```\n90 AIME seeds → apply 4 instruction wrappers (+ reward variants) →\nquery teacher (DeepSeek‑R1) → collect reasoning traces → fine‑tune Qwen2.5‑Instruct\n```\n\n---\n\n## Reproduce\n\n```bash\n# 1) Build wrapped prompts from seeds\npython DataConstruction/build_prompt_variants.py \\\n  --input DataConstruction/seeds.csv \\\n  --out DataConstruction/variants.csv\n\n# 2) Query teacher model to collect reasoning traces\npython DataConstruction/deepseek_query.py \\\n  --input DataConstruction/variants.csv \\\n  --out DataConstruction/DS_responses.csv\n\n# 3) Combine Full Data\npython DataConstruction/combine_deepseek_data.py\n```\n\n---\n\n## Citation\n```\n@article{bsharat2025prompting,\n  title={Prompting Test-Time Scaling Is A Strong LLM Reasoning Data Augmentation},\n  author={Bsharat, Sondos Mahmoud and Shen, Zhiqiang},\n  journal={arXiv preprint arXiv:2510.09599},\n  year={2025}\n}\n```\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fvila-lab%2Fptts","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fvila-lab%2Fptts","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fvila-lab%2Fptts/lists"}