SamahFodeh/EPPC-CARE-Qwen2.5-32B

EPPC-CARE: Qwen2.5-32B

Distribution of our best reported EPPC-CARE model: the Qwen2.5-32B LoRA adapter and the code that runs its multistage annotation workflow. The package is approximately 527 MiB. Base-model weights are downloaded and cached automatically on first use.

Setup

Use Linux, Python 3.11 or 3.12, and NVIDIA CUDA GPUs. The 16-bit base model requires approximately 65 GB for weights alone, plus inference and retrieval memory. Use sufficient GPU capacity, such as a 96 GB GPU or multiple GPUs. For an 80 GB GPU, --gpu-memory-utilization 0.9 --retrieval-device cpu may be needed; the exact fit depends on the software stack and available memory.

This repository is public; no Hugging Face login is required to download it.

python -m venv .venv
source .venv/bin/activate
pip install huggingface_hub
hf download SamahFodeh/EPPC-CARE-Qwen2.5-32B --local-dir eppc-care
cd eppc-care
pip install -r requirements.txt

Retrieval data

The current runner requires the original 607 labeled training examples as retrieval memory. Obtain eppc_train.jsonl through the group's existing data sharing process; it is separate from this model repository. Preserve row order when reproducing our reported evaluation. No test data is needed to annotate new messages.

--retrieval-data accepts JSONL or a datasets.save_to_disk() directory. Each training record needs string query and answer fields. answer contains JSON with a results list of objects containing Code, Sub-code, and Span. The runner checks JSONL against the original training-file checksum.

Annotate one message

python run.py \
  --retrieval-data /path/to/eppc_train.jsonl \
  --message "Thank you for explaining the next steps. When should I schedule the follow-up?" \
  --to-patient N \
  --output predictions.jsonl

Y means provider-to-patient; N means patient-to-provider.

The runner uses BGE to retrieve two labeled training examples, then applies the adapter through all three generation stages:

1. Extract communication-intent spans.
2. Refine spans and assign parent Codes.
3. Assign a Sub-code to each span/Code pair using the valid child names.

The runner loads unsloth/Qwen2.5-32B-Instruct and attaches our adapter through vLLM. No manual merge is needed. A plain transformers.pipeline() call does not run the multistage controller.

Batch input and output

For batch inference, create a JSONL file with one object per line:

{"message": "Thank you for explaining the next steps.", "to_pat_yn": "N"}
{"message": "Please schedule a follow-up appointment next week.", "to_pat_yn": "Y"}

Then run:

python run.py \
  --retrieval-data /path/to/eppc_train.jsonl \
  --input messages.jsonl \
  --output predictions_batch.jsonl

Original dataset records with query and optional answer fields are also accepted, as are datasets.save_to_disk() directories. Queries must contain TO_PAT_YN: Y or TO_PAT_YN: N.

Each output row contains query, stage1_out, stage2_out, and final_out. final_out is a JSON string with the final results list. If the input includes answer, the output also includes gold_answer. Existing output files are not overwritten.

Useful options:

• --tensor-parallel-size 2: spread the base model across two GPUs.
• --retrieval-device cpu: run the BGE retrieval model on CPU.
• --dry-run: validate local inputs without downloading models or using a GPU.
• --adapter: use another local adapter directory or a Hugging Face model ID.
• --base-revision / --adapter-revision: select explicit Hub revisions.
• --merged-model: use locally merged weights instead of runtime LoRA.

Run python run.py --help for the full argument list.

Evaluation

With the original labeled 260-case test file:

python run.py --retrieval-data /path/to/eppc_train.jsonl \
  --input /path/to/eppc_test.jsonl --output predictions_test.jsonl
python score.py predictions_test.jsonl

Metric	Reported micro-F1
Relaxed Span	86.083%
Relaxed Code+Span	70.495%
Relaxed Sub-code+Span	67.038%
Relaxed Triplet	64.401%
Code only	87.061%
Sub-code only	80.508%
Code+Sub-code only	77.332%

These saved results used locally merged 16-bit weights and the original retrieval data. Runtime LoRA uses the same stage prompts and orchestration but may produce different numerical outputs. The supplied controller passed CPU checks, including prompt parity across all 260 cases; a fresh GPU evaluation of the runtime-LoRA path has not been performed.

score.py computes the four span-dependent endpoints. evaluation.json contains the saved aggregate results. validation.json records packaging checks.

For the original merged-weight path, install requirements-train.txt in a separate environment and run python merge.py --adapter . --output merged_model. Then return to the vLLM environment and add --merged-model merged_model to the inference command. eval_original.py is the unchanged original evaluator.

The repository contains no training or test message records. The examples under examples/ are synthetic.