Instructions to use amitaddi/dronam-v1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use amitaddi/dronam-v1 with PEFT:
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("unsloth/gemma-4-e2b-it-unsloth-bnb-4bit") model = PeftModel.from_pretrained(base_model, "amitaddi/dronam-v1") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- Unsloth Desktop
Dronam-V1 β orthopaedic trauma case report β structured case record
Dronam-V1 is a LoRA adapter (rank 16) on Gemma-4-E2B-it that turns an English orthopaedic trauma case description into a 30-field JSON case record (bone, segment, side, open fracture and Gustilo grade, neurovascular deficit, compartment syndrome, classification as stated, imaging, temporising and definitive treatment, implants, outcome). It is built to pre-fill a planning record for a surgeon to check, running locally on one consumer GPU.
It is a drafting aid, not a medical device. It still fills about one in four fields that the source never states (see below). Every field must be confirmed by a clinician before use.
Results on held-out data
200 case reports from articles never seen in training (split by article), schema-constrained decoding, one RTX 5060 Ti 16 GB. Reference labels are a three-model majority vote (not yet surgeon-reviewed), so all numbers are provisional. 95% bootstrap CIs.
| Model | Macro field score | Hallucination β | Safety-field hallucination β | Schema-valid | s / case |
|---|---|---|---|---|---|
| Dronam-V1 | 77.1% (76.0β78.1) | 24.7% | 7.6% | 100% | 2.0 |
| Dronam-V1, no constraint | 77.0% | 24.7% | 7.6% | 96.0% | 2.0 |
| Gemma-4-E2B (base) | 70.9% (69.7β72.1) | 27.3% | 20.2% | 97.0% | 2.6 |
| Gemma-4-E4B | 69.6% | 31.1% | 34.8% | 98.0% | 4.5 |
| MedGemma-4B-it | 59.7% | 52.3% | 46.8% | 93.0% | 4.4 |
| Nemotron-3-Nano-4B | 66.6% | 37.1% | 21.9% | 100% | 3.3 |
| Bonsai-27B | 70.5% | 21.0% | 25.5% | 98.5% | 8.5 |
Paired vs the base: field score +6.2 points (95% CI 5.1β7.3); safety-field hallucination β15.5 to β9.7
points. Teacher models that produced the labels (Gemma-4-12B 83.7%, Qwen3.5-9B 82.6%) score higher but
are inflated by construction; the full table (12 runs, 11 models) and per-field results are in the
paper and in eval/heldout_silver/ of the project repository.
Safety fields: side, open fracture, Gustilo grade, neurovascular deficit, compartment syndrome, stated AO/OTA code. Hallucination: the reference says "not stated" and the model filled a value.
Earlier numbers withdrawn. A previous version of this card quoted a 30-case "silver pilot" (0.765 field score) and cloud-model comparisons. 24 of those 30 cases were in the training data, and the cloud figures were not measured. Both are withdrawn; use the held-out results above.
How to use
Send exactly the instruction in instruction.txt followed by the case text, and constrain the output to
case_record.schema.json if your server supports JSON-schema structured output (llama.cpp, LM Studio,
vLLM do).
from unsloth import FastModel # or transformers + peft
model, tok = FastModel.from_pretrained("amitaddi/dronam-v1", load_in_4bit=True)
instruction = open("instruction.txt").read()
msgs = [{"role": "user", "content": instruction + "\n\nTEXT:\n" + case_text.strip()}] # exactly as in training
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt", return_dict=True).to("cuda")
out = model.generate(**ids, max_new_tokens=1200, do_sample=False)
print(tok.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=True))
A merged 4-bit GGUF for llama.cpp / LM Studio is in amitaddi/dronam-v1-GGUF.
Training
- Data: 2,470 English orthopaedic trauma case reports (CC BY / CC0; MultiCaRe and Europe PMC), labelled by gpt-oss-20B, Gemma-4-12B and Qwen3.5-9B with field-wise majority vote. Validation 232, test 200, split by source article.
- QLoRA 4-bit, LoRA r = 16, Ξ± = 16 on decoder attention and MLP; 2 epochs, lr 2e-4 cosine, effective batch 8, loss on the JSON answer only. Validation loss 0.128 β 0.121.
Limitations
Silver (model-made) reference labels; published case reports, not hospital notes; English only; hallucination rate far above what unsupervised auto-fill needs; regresses on some fields (3-D CT reconstruction, associated injuries, bone graft). Do not use on identifiable patient data outside an approved, local deployment.
Licence
Apache-2.0 (as the Gemma-4 base). Training texts are CC BY / CC0 with per-article attribution in the project repository.
- Downloads last month
- 12