How kaistai/cot-collection works
Project Type and Purpose
• This is a research‐oriented pipeline for Chain-of-Thought (CoT) data augmentation and fine-tuning of T5 models.
• It consists of two primary phases:
a) Rationale Augmentation – generate synthetic rationales via OpenAI APIs
b) Fine-tuning – train CoT-enhanced T5 models (3B & 11B) using DeepSpeed and Huggingface.
Key Components (to represent
A. Data Sources Layer
1. Original CoT Collection
– Huggingface dataset ("kaist-ai/CoT-Collection")
2. Demonstration Prompts
– Plain-text prompt files under
CoT_Rationale_Augmentation/demonstration/{task}/{prompt.txt}
B. Rationale Augmentation Module
1. Orchestration Scripts (shell)
– scripts/{flan.sh, sni.sh, t0.sh, t0p.sh, additional.sh}
2. Python Runner & Utils
– CoT_Rationale_Augmentation/run.py
– CoT_Rationale_Augmentation/utils.py
3. External Service
– OpenAI API (keys in api_keys.json → api.json)
4. Outputs
– Generated rationales saved in
CoT_Rationale_Augmentation/outputs/{subset}/{model_name}/rat/temp_{temperature}/
C. Data Extraction & Merging
1. Data Directory
– data_extraction/data/{split}/{subset}/
2. Extraction Logic
– Merges original + augmented into
codex_rationale_{split}_{phase}.json
D. Fine-Tuning Module
1. Configuration Files
– CoT_Finetuning/configs/train/*.json
– CoT_Finetuning/configs/eval/*.json
2. Training Code
– CoT_Finetuning/src/run.py
– CoT_Finetuning/src/run_peft.py
– CoT_Finetuning/src/T5.py, Datasets_end2end.py, evaluation.py, utils.py
3. DeepSpeed Engine
– CoT_Finetuning/src/deepspeed.json
4. External Libraries
– Huggingface Transformers & Datasets
5. Outputs
– Checkpoints: “kaist-ai/CoT-T5-3B” and “kaist-ai/CoT-T5-11B”
E. Model Serving / Usage
– Example in README: loading via
transformers.AutoModelForSeq2SeqLM.from_pretrained(...)
Relationships & Data Flow
1. Original Dataset + Demonstrations → (Rationale Augmentation Scripts & OpenAI API) → Augmented Rationales
2. Augmented Rationales → (Data Extraction) → Merged JSON ready for training
3. Merged JSON + Training Configs → (CoT_Finetuning with DeepSpeed & Transformers) → Fine-tuned Model
4. Fine-tuned Model → (Huggingface Hub) → End-user loading via Transformers
Architectural Patterns & Principles
• Pipeline/ETL structure: clear separation of augmentation, extraction, and training phases
• Modularity: each phase in its own top-level directory
• Configuration-driven: JSON configs for training & evaluation
• External dependencies managed via scripts and api_keys.json