How duxiaoman-di/xuanyuan works
XuanYuan is a Chinese finance-focused language-model project: it publishes model families, interactive inference, a FinanceIQ benchmark, and a DeepSpeed training toolkit. The user-facing paths are either asking questions in the chat demo and receiving generated text, or evaluating a model/API against finance exam questions and saved accuracy results. The README also documents FinX1’s thinking, reward-model, and reinforcement-learning approach; those training stages are documented capabilities, not sampled implementation modules.
Main entry points: utils.py Evaluation runner · cli_demo.py Chat demo · model_hook.py Model and data hooks · dataset.py Training datasets
Components 19
Chat inference
- Chat demo
cli_demo.py - Conversation templates
conversation.py - Language model
FinanceIQ evaluation
- FinanceIQ questions
FinanceIQ/data - Evaluation runner
FinanceIQ/src/utils.py - Exam subjects
FinanceIQ/src/categories.py - Local model evaluator
FinanceIQ/src/hf_causal_model.py - ChatGPT evaluator
FinanceIQ/src/chatgpt.py - GPT-4 evaluator
FinanceIQ/src/gpt4.py - ERNIE evaluator
FinanceIQ/src/ernie.py - Evaluation results
FinanceIQ/src/utils.py
Model training
- Training orchestrator
llm-code/dxm_llm_main.py - Model and data hooks
llm-code/model_hook.py - Training datasets
llm-code/dataset.py - Training configuration
llm-code/config.py - Pretraining data
llm-code/pretrain_data_process.py - FinX1 reasoning approach
Other components
- User
- Evaluator
Connections 18
- User → Chat demo: asks
cli_demo.py - Chat demo → Conversation templates: formats turns
cli_demo.py - Chat demo → Language model: generates text
cli_demo.py - Evaluator → Evaluation runner: starts evaluation
FinanceIQ/src/utils.py - Evaluation runner → Exam subjects: iterates subjects
FinanceIQ/src/utils.py - Evaluation runner → FinanceIQ questions: loads questions
- Evaluation runner → Local model evaluator: runs evaluator
FinanceIQ/src/hf_causal_model.py - Evaluation runner → ChatGPT evaluator: runs evaluator
FinanceIQ/src/chatgpt.py - Evaluation runner → GPT-4 evaluator: runs evaluator
FinanceIQ/src/gpt4.py - Evaluation runner → ERNIE evaluator: runs evaluator
FinanceIQ/src/ernie.py - Local model evaluator → Language model: loads checkpoint
FinanceIQ/src/hf_causal_model.py - Evaluation runner → Evaluation results: saves scores
FinanceIQ/src/utils.py - Training orchestrator → Model and data hooks: gets training hooks
llm-code/dxm_llm_main.py - Model and data hooks → Training datasets: builds dataloader
llm-code/model_hook.py - Model and data hooks → Training configuration: gets settings
llm-code/model_hook.py - Training datasets → Training orchestrator: logs loading
llm-code/dataset.py - Pretraining data → Training datasets: prepares corpus
- FinX1 reasoning approach → Language model: guides training