---
license: llama3
language:
- en
library_name: peft
pipeline_tag: text-generation
base_model: vineetdaniels/NYXMed-V17-Merged
tags:
- medical
- radiology
- medical-coding
- icd-10
- cpt
- llama-3
- lora
- peft
- healthcare
---

# NYXMed V18 — Radiology Coding LoRA Adapter

LoRA adapter trained on top of [`vineetdaniels/NYXMed-V17-Merged`](https://huggingface.co/vineetdaniels/NYXMed-V17-Merged),
targeting **primary-ICD accuracy** with proximity-ranked retrieval candidates.

For a deployable single model, use [`vineetdaniels/NYXMed-V18-Merged`](https://huggingface.co/vineetdaniels/NYXMed-V18-Merged).

## Highlights
- **Best eval_loss: 0.0710** (early-stopped at step 1,700; best checkpoint step 1,400)
- Trained on **59,170** coder-verified examples weighted toward primary-ICD corrections (family-swaps 47%)
- Built on the **proximity-ranking retrieval fix** (+21.5pp recall@10 of the correct primary on
  previously-wrong records) — **must be deployed with the matching preprocessor change**

## Training
| | |
|---|---|
| Base | `vineetdaniels/NYXMed-V17-Merged` |
| LoRA | r=64, α=128, dropout=0.05, targets q/k/v/o/gate/up/down_proj |
| Examples | 59,170 (weighted) | 
| Effective batch | 32 | LR | 1e-5 cosine | Max len | 2,560 |
| Hardware | 4× H200, ~10.9h | Attn | sdpa | DeepSpeed | ZeRO-3 |

## Usage
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch
base = AutoModelForCausalLM.from_pretrained("vineetdaniels/NYXMed-V17-Merged", torch_dtype=torch.bfloat16, device_map="auto")
tok  = AutoTokenizer.from_pretrained("vineetdaniels/NYXMed-V18-Model")
model = PeftModel.from_pretrained(base, "vineetdaniels/NYXMed-V18-Model").eval()
```

eval_loss is on V18's own held-out split (not directly comparable to V17's split). The authoritative
metric is primary-ICD accuracy on a common held-out production set. Radiology-only, review-then-accept use.