File size: 9,279 Bytes
7c112bd 4d6826d 52768f3 4d6826d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 | ---
license: apache-2.0
---
<p align="center">Clinical + AI β An Open-Source Library for Clinical Diagnosis Support</p>
A GPU-based LLM fine-tuning and serving pipeline. This document describes the project's **LLM model composition** from the perspectives of base model, training, data, and inference/serving.
---
## 1. Model Composition Overview
| Category | Details |
|------|------|
| Base model | **Llama 3.1 8B Instruct** |
| Extended pipeline | **Qwen3-32B** multi-phase (SFT β DPO β GRPO β Export) |
| Fine-tuning method | LoRA (PEFT) |
| Deployed models | LoRA adapter / base+LoRA merged model |
| Distribution channel | HuggingFace Hub β `the-platforms/MediCPX` |
| Domain | CPX medical (clinical diagnosis) instruction-following |
**Two pipeline tracks**
- **Llama single pipeline** β `FineTuningPipeline`: preprocessing β tokenization β LoRA training β save
- **Qwen3 multi-phase pipeline** β `MultiPhaseFineTuningPipeline`: SFT β DPO β GRPO β Export
---
## 2. Base Model & Weights
At the end of training, artifacts are produced under a per-run timestamped directory.
```
{output_dir}/{yyyymmddhhmm}/final_model/
βββ adapter/ # LoRA adapter (tens of MB) β for shared-base serving
β βββ adapter_config.json
β βββ adapter_model.safetensors
βββ merged/ # base + LoRA merged model (tens of GB) β for standalone deployment
β βββ config.json
β βββ model.safetensors
βββ model_info.json # base model, training config, and path metadata
```
| Format | Size | Loading | Use |
|------|------|------|------|
| `adapter/` | Small | `PeftModel.from_pretrained(base, adapter)` | Experiments & versioning |
| `merged/` | Large | `AutoModelForCausalLM.from_pretrained(merged)` | Dependency-free standalone serving |
**HuggingFace Hub upload** β `runs/upload_to_hf.py`
```bash
export HF_TOKEN=hf_xxx
python runs/upload_to_hf.py --folder-path /ai_models/merged --repo-id the-platforms/MediCPX
```
- Training-only state (`optimizer.pt`, `scheduler.pt`, `rng_state*`) is excluded by default
- Details: [`llama/huggingface_upload.md`](./llama/huggingface_upload.md)
---
## 3. Training Configuration
**4-step flow**: preprocessing β tokenization β LoRA training β save
**Key hyperparameters**
| Item | Value |
|------|-----|
| LoRA rank | `r = 8` (Ξ± = r Γ 2) |
| Learning rate | `2e-4` |
| max_length | `2048` |
| Precision | bf16 (auto) |
| Other | Response-region loss masking + sequence packing |
**Key code paths**
| Category | Path |
|------|------|
| Pipeline package | `app/finetuning/` |
| Llama single pipeline | `app/finetuning/pipeline.py` (`FineTuningPipeline`) |
| Qwen3 multi-phase | `app/finetuning/pipeline.py` (`MultiPhaseFineTuningPipeline`) |
| LoRA trainer | `app/finetuning/training/llama_trainer.py` |
| GRPO trainer (TRL) | `app/finetuning/training/grpo_trainer.py` |
| Reward functions | `app/finetuning/training/rewards/` (`AgentReward`, `SIDomainReward`) |
| LLaMA-Factory integration | `app/finetuning/llamafactory/` |
**Run**
```bash
# Llama LoRA fine-tuning
python runs/run_finetuning.py --model-path /ai_models --data-path /data/output/1105 --epochs 3
# Multi-GPU (DDP)
torchrun --nproc_per_node=2 runs/run_finetuning.py --model-path /ai_models --data-path /data/output/1105
# Qwen3 multi-phase
python runs/run_qwen3_finetuning.py
```
Details: [`llama/process.md`](./llama/process.md) Β· [`qwen3_finetuning_process.md`](./qwen3_finetuning_process.md)
---
## 4. Dataset Configuration
PDF source documents β instruction-response JSON generation (4-Stage).
```
PDFLoader(PyMuPDF4LLM) β MarkdownConverter β InstructionParser(GPT-4) β BatchedJSONWriter
```
**Output schema** (identical to training input)
```json
[{ "instruction": "...", "context": "(optional)", "response": "..." }]
```
**Data composition by training method**
| Method | Format | Fields |
|------|------|------|
| SFT | ShareGPT `messages` | system/user/assistant β one gold answer |
| DPO | `sharegpt_dpo` (ranking) | `conversations` + `chosen` + `rejected` |
| GRPO | Prompt + reward function | No gold answers or pairs needed; scored via `RewardFunction.compute()` |
**Code paths**
| Category | Path |
|------|------|
| Preprocessing package | `app/data_handling/data_pre_processing/` |
| Qwen3 data pipeline | `app/data_handling/` (chunker, qa_generation, quality, ragas_eval, trajectory) |
| Entry points | `runs/run_cpx_processing.py`, `runs/run_qwen3_data_pipeline.py` |
```bash
INPUT_PATH=/data/input/cpx.pdf OUTPUT_PATH=/data/output/1105 python runs/run_cpx_processing.py
```
Details: [`llama/data_preprocessing.md`](./llama/data_preprocessing.md)
---
## 5. Inference & Serving
| Category | Path | Description |
|------|------|------|
| Lightweight inference server | `server.py` | Single-model startup; `/generate`, `/stream`, `/health` |
| Full pipeline API | `main.py` | Integrated model load/training/inference; OpenAI-compatible chat |
| vLLM serving | `app/serving/vllm_config.py` | For Qwen3 Agent; Hermes tool parser |
| Model load helper | `runs/load_llama_3_1_8b.py` | Llama 3.1 8B load verification |
```bash
# Lightweight inference server
MODEL_DIR=/path/to/model USE_4BIT=1 uvicorn server:app --host 0.0.0.0 --port 8080
# vLLM (with tool calling)
./scripts/serve_qwen3_vllm.sh
```
**OpenAI-compatible REST API** β two entry points
`main.py` β Full Pipeline API (`:8000`)
| Category | Endpoints |
|------|-----------|
| Health | `GET /health` |
| Model | `POST /api/v1/models/load`, `GET /api/v1/models/{id}/verify` |
| Preprocess | `POST /api/v1/preprocessing/run` |
| Tokenize | `POST /api/v1/tokenization/tokenize` |
| Training | `POST /api/v1/training/start`, `GET /api/v1/training/{job_id}/status` |
| Multi-Phase | `POST /api/v1/training/multi-phase/start` |
| Data Pipeline | `POST /api/v1/data-pipeline/generate-qa` |
| Inference | `POST /api/v1/inference/load`, `GET /api/v1/inference/models` |
| Chat (OpenAI-compatible) | `POST /api/v1/chat/completions`, `POST /v1/chat/completions` |
`server.py` β Inference API (`:8080`): `POST /generate`, `POST /stream` (SSE), `GET /health`
```bash
./scripts/api_server.sh start # start main.py
```
Request examples: [`API_SAMPLE.md`](./API_SAMPLE.md)
---
## 6. Evaluation
| Category | Path |
|------|------|
| Agent evaluation | `runs/eval_agent.py` |
| Domain evaluation | `runs/eval_domain.py` |
| Model evaluation | `runs/eval_model.py` |
| NLG surface metrics (BLEU/ROUGE/METEOR/BERTScore) | `runs/eval_nlg.py` |
| RAGAS evaluation | `runs/eval_ragas.py` |
| Model comparison | `runs/eval_compare.py` |
Details: [`evaluation.md`](./evaluation.md) Β· [`llama/evaluation_usage.md`](./llama/evaluation_usage.md) Β· [`llama/evaluation_runbook.md`](./llama/evaluation_runbook.md)
---
## 7. Demo
| Category | Path | Description |
|------|------|------|
| Medical chat loop demo | `scripts/medical_chat_loop.sh` | Calls the chat API in parallel with N random medical questions (repeats periodically) |
| Question set | `scripts/medical_questions.txt` | Question pool for demo input |
| System prompt | `scripts/medical_system_prompt.txt` | System prompt for the demo |
| Frontend integration guide | [`frontend_request_cancel_guide.md`](./frontend_request_cancel_guide.md) | Request/cancel integration |
```bash
# After starting the inference server
API_URL=http://localhost:8000 MODEL=MediCPX ./scripts/medical_chat_loop.sh
```
---
## Appendix β Deliverables Inventory
| # | Item | Key Location | Form |
|---|------|-----------|------|
| 1 | Model weights | `final_model/`, HF Hub `the-platforms/MediCPX` | LoRA adapter / merged model |
| 2 | Training code | `app/finetuning/`, `runs/run_finetuning.py` | Python package + entry point |
| 3 | Inference code | `server.py`, `main.py`, `app/serving/` | FastAPI + vLLM |
| 4 | Dataset | `app/data_handling/`, `runs/run_cpx_processing.py` | PDF β instruction JSON |
| 5 | Demo | `scripts/medical_chat_loop.sh` | CLI chat demo |
| 6 | API | `main.py` (Full) / `server.py` (Inference) | OpenAI-compatible REST |
| 7 | Technical docs | `docs/` | Markdown + papers/patents |
**Technical documentation**
| Category | Document |
|------|------|
| Project overview | [`README.md`](./README.md), [`QUICKSTART.md`](./QUICKSTART.md) |
| End-to-end summary | [`llama/end_to_end_summary.md`](./llama/end_to_end_summary.md) |
| Data preprocessing | [`llama/data_preprocessing.md`](./llama/data_preprocessing.md) |
| Training walkthrough | [`llama/process.md`](./llama/process.md) |
| Qwen3 multi-phase | [`qwen3_finetuning_process.md`](./qwen3_finetuning_process.md) |
| Evaluation | [`evaluation.md`](./evaluation.md), [`llama/evaluation_usage.md`](./llama/evaluation_usage.md), [`llama/evaluation_runbook.md`](./llama/evaluation_runbook.md) |
| Improvement history | [`llama/improvements.md`](./llama/improvements.md) |
| Research & papers | [`llama/research.md`](./llama/research.md), [`llama/papers.md`](./llama/papers.md) |
| HF upload | [`llama/huggingface_upload.md`](./llama/huggingface_upload.md) |
| Deployment | [`deployments/README.md`](./deployments/README.md) (Docker, K3s) |
| Patent/paper drafts | `docs/paper/` (CPX medical-domain IMRaD, patent application drafts & figures) |
|