feat: add MiniMax as configurable evaluation LLM provider (default M3)

[octo-patch]

Summary

• Add configurable evaluation LLM client (pipeline/benchmarks/utils/eval_llm.py) supporting OpenAI and MiniMax providers with auto-detection, temperature clamping, and think-tag stripping
• Default MiniMax model is MiniMax-M3 (512K context, 128K max output, image input support); MiniMax-M2.7 and MiniMax-M2.7-highspeed remain available
• Update MagnifierBench, MathVista, and MM-Vet evaluation datasets to use the configurable client instead of hardcoded OpenAI API calls, with backward-compatible eval_provider parameter
• Update Syphus data generation pipeline with MiniMax provider documentation and temperature handling
• Add 24 unit tests and 4 integration tests (all passing)

Motivation

The benchmark evaluation system (MagnifierBench, MathVista, MM-Vet) previously hardcoded OpenAI GPT-4 as the evaluation judge LLM. This PR makes the evaluation LLM configurable, enabling users to choose alternative providers like MiniMax M3 — the latest MiniMax model with 512K context and image input support — as a cost-effective evaluation backend.

Configuration

# For benchmark evaluation (defaults to MiniMax-M3)
export EVAL_LLM_PROVIDER="minimax"
export MINIMAX_API_KEY="your-key"

# For Syphus data generation (via liteLLM)
export MINIMAX_API_KEY="your-key"
export OPENAI_API_ENGINE="openai/MiniMax-M3"
export OPENAI_API_BASE="https://api.minimax.io/v1"

Or via YAML config:

datasets:
  - name: magnifierbench
    eval_provider: minimax
    api_key: your-key

To target a specific MiniMax model instead of the default:

client = EvalLLMClient(provider="minimax", api_key="your-key", model="MiniMax-M2.7")

Changes

File	Change
pipeline/benchmarks/utils/eval_llm.py	New configurable LLM client with provider registry; default MiniMax model is M3
pipeline/benchmarks/utils/init.py	Package init
pipeline/benchmarks/datasets/magnifierbench.py	Use EvalLLMClient, add eval_provider/eval_model params
pipeline/benchmarks/datasets/mathvista.py	Use EvalLLMClient, add eval_provider/eval_model params
pipeline/benchmarks/datasets/mmvet.py	Use EvalLLMClient, replace OpenAI() client
mimic-it/syphus/file_utils.py	Add MiniMax docs, temp clamping, query_llm() alias
unit_tests/test_eval_llm.py	24 unit tests
unit_tests/test_eval_llm_integration.py	4 integration tests
README.md	MiniMax badge, config docs

MiniMax model line

Model	Notes
MiniMax-M3	Default. 512K context, 128K max output, image input support
MiniMax-M2.7	Previous-generation model
MiniMax-M2.7-highspeed	Previous-generation low-latency variant

Test plan

• 24 unit tests passing (provider config, init, temp clamping, think-tag stripping, chat completion, retry logic)
• 4 integration tests passing against live MiniMax API (basic completion, judge yes/no, scoring, auto-detect)
• Verify backward compatibility: existing OpenAI-based evaluation works unchanged when no eval_provider is set