prerequisites.md

Prerequisites

Hardware

Target model	Params	VRAM required	Minimum GPU
Gemma 3 27B-IT	27B	~55 GB	1×A100 80 GB
Mistral Small 3.1 24B	24B	~48 GB	1×A100 80 GB
Qwen 2.5 32B-IT	32B	~64 GB	1×A100 80 GB
Llama 3.1 70B-IT	70B	~140 GB	2×H100 80 GB
Qwen3-235B-A22B (data gen only)	235B MoE	~240 GB (fp8)	2×H200

Software

• Python 3.10+
• uv

Accounts

• HuggingFace account with approved access to:
  • google/gemma-3-27b-it
  • meta-llama/Llama-3.1-70B-Instruct
  • meta-llama/Llama-3.2-1B-Instruct (low-resource path)
  • meta-llama/Llama-3.2-3B-Instruct (low-resource path)
  • google/gemma-2-2b-it (low-resource path)
• HF_TOKEN environment variable set to your token
• OPENAI_API_KEY environment variable set to your key (required for three-phase labeling in Phase 3)