Llama-rasopathy-vus-tier1-scout-adapter

Built with Llama

A LoRA adapter fine-tuned on top of Llama-4-Scout-17B-16E-Instruct to generate structured, source-grounded prioritization reports for Tier 1 Variants of Uncertain Significance (VUS) in RASopathy-associated genes. Trained via the Adaption AutoScientist platform for the Adaption AutoScientist Challenge 2026 (Science category).

This adapter corresponds to the Rephrase OFF configuration of a paired experiment (see "Comparison context" below).

---

What it does

Given the computational evidence for a variant — CADD PHRED score, investigation score, concordant predictor count, and ClinGen/GenCC gene-disease validity classification — the adapter produces a technical report that:

• Reports only the computational values present in the input.
• States the ClinGen GCEP validity classification and condition match type.
• Always includes the disclaimer that Tier 1 prioritization is a computational investigation priority, not an ACMG/AMP clinical reclassification.
• Notes that functional evidence, family segregation data, and expert-panel review are required before any formal reclassification.

It is a research/tooling aid for variant triage. It does not classify variants clinically and must not be used as a diagnostic device.

---

Intended use and limitations

Intended: computational triage support for clinical-genomics researchers working on RASopathy variant curation; a demonstration of source-grounded report generation.

Not intended: clinical diagnosis, variant reclassification, patient management, or any use where the output is treated as a medical determination. Outputs are model-generated and can contain errors, including plausible-looking but unsourced statements. Every output should be verified against primary sources (ClinVar, ClinGen, GenCC) before any downstream use.

Known failure mode: LLM-based generation can introduce content not present in the input (e.g. cDNA/HGVS-c notation, MONDO identifiers, or claims of functional/segregation evidence). This was the central subject of the audit that produced this adapter; the training data and blueprint were specifically designed to reduce — not eliminate — that behavior. Do not assume the output is fully traceable to the input without checking.

---

Training details

• Base model: meta-llama/Llama-4-Scout-17B-16E-Instruct (109B total, Mixture-of-Experts)
• Method: Supervised fine-tuning (SFT), LoRA
• Data format: chat (train_on_inputs = false — loss on the assistant turn only)

Hyperparameter	Value
LoRA rank (r)	64
LoRA alpha	128
LoRA dropout	0
Epochs	4
Learning rate	1e-4
LR scheduler	cosine (0.5 cycles)
Warmup ratio	0.05
Weight decay	0.02
Max grad norm	1
Trainable modules	k_proj, o_proj, q_proj, v_proj, shared_expert.{gate,up,down}_proj, feed_forward.{gate,up,down}_proj

Training dataset: Fernandosr85/adaption-rasopathy-vus-tier1-reports-v1 — 163 variants across 10 RASopathy genes (SOS1, PTPN11, RAF1, BRAF, MAP2K1, MAP2K2, RIT1, SHOC2, SOS2, HRAS). All required facts are drawn from ClinGen (CC BY 4.0) and GenCC (CC0 1.0) snapshots dated 2026-06-28. Reference: Wilcox EH, DiStefano MT et al. 2025, PMID:40496714.

Training target: the deterministic output column (template-built, fully source-traceable) — not the platform-generated enhanced_completion column. The enhanced completions are included in the dataset for audit/comparison only, and were deliberately excluded from training after they were found to introduce unsourced content.

---

Comparison context (important for honest interpretation)

This adapter is one half of a paired experiment on the Adaption platform. The two halves differ in three variables at once, not one:

	This adapter (OFF)	Paired adapter (ON)
Adaption "Prompt Rephrase"	OFF	ON
Base model	Llama-4-Scout-17B	Mixtral-8x7B
LoRA rank / epochs	64 / 4	16 / 2

Because base model and hyperparameters changed alongside the Rephrase toggle, any observed difference in output quality cannot be attributed to the Rephrase setting alone. The paired comparison shows the difference between the two final adapters, not an isolated ablation of one setting. A companion Space visualizes this comparison on real published data: rasopathy-rephrase-on-vs-off.

---

How to use

This is a LoRA adapter — load it on top of the base model with PEFT:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE = "meta-llama/Llama-4-Scout-17B-16E-Instruct"
ADAPTER = "Fernandosr85/Llama-rasopathy-vus-tier1-scout-adapter"

tokenizer = AutoTokenizer.from_pretrained(BASE)
base = AutoModelForCausalLM.from_pretrained(BASE, device_map="auto", torch_dtype=torch.bfloat16)
model = PeftModel.from_pretrained(base, ADAPTER)
model.eval()

messages = [{"role": "user", "content": "<your variant prioritization prompt>"}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=700, do_sample=False)  # greedy = deterministic
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Running the base Llama-4-Scout (109B MoE) needs a large GPU (A100 80GB recommended, or 4-bit quantization).

---

License and attribution

This model is a derivative of Llama 4 and is distributed under the Llama 4 Community License Agreement.

Llama 4 is licensed under the Llama 4 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.

Full license text: https://github.com/meta-llama/llama-models/blob/main/models/llama4/LICENSE

Per the license, use of this adapter requires the base Llama 4 model, which is subject to Meta's Acceptable Use Policy. If your product or service exceeded 700 million monthly active users on the Llama 4 release date, you must request a separate license from Meta.

The fine-tuning modifications in this adapter are owned by the adapter author; the underlying Llama 4 materials remain owned by Meta.

This model card is provided for research documentation. It is not legal advice; consult the license text directly for authoritative terms.

---

Citation

If you use this adapter, please reference the training dataset and the underlying curation source:

Wilcox EH, DiStefano MT, et al. (2025). PMID:40496714,
DOI:10.1016/j.gimo.2025.103430.