patientagent-cft-augmented-sft-r16

PatientAgent CFT-augmented SFT (rank 16)

This repository contains one experimental LoRA adapter for research on simulated patient responses in clinical dialogue. It is not a medical device, must not be used for patient care, and must not be treated as a source of medical advice.

What this model is

The adapter is an application of the training idea in Teaching LLMs How to Learn with Contextual Fine-Tuning to PatientAgent's conditional response task. During training, one of ten generic learning-strategy prompts was sampled uniformly and prepended to each canonical SFT context. Loss was still computed only on the exact original patient-response target tokens. The contextual prompt is absent at inference.

This is therefore CFT-augmented SFT, not an exact reproduction of the paper's raw-corpus CFT experiment. Pure raw-text CFT would change PatientAgent's conditional task and train on clinician turns, so it was not used. The implementation is clean-room; no source from the authors' repository was copied. The official reference implementation was reviewed at commit 7b709820c091b40e3e6153fe0a6dceccef75417d; that repository did not declare a top-level license at the reviewed commit.

Parent and training data

• Base model: Qwen/Qwen3.5-4B, resolved training snapshot 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a (Apache-2.0).
• Dialogue source: MTS-Dialog training CSV at commit 3ff0801933608d6f570468c13125125fb5cabdea, SHA-256 65a28681dd59fc159681ea026e44610f2af7bc64a0e7f892d763eb03d9f503dc (CC BY 4.0).
• Case facts: cs552-the-expendables/mts-rl-training-data at revision f188dfcb71fc13a74446b5ad34df9369f1534e57, SHA-256 cf85d176fd64c8575e08feb7e94a63b0baa6aeff0d6e90a28d5cc25d4fe0248e.
• Scope: the same 1,200 fact-backed training dialogues and 4,877 patient-turn targets used by the canonical current SFT pipeline. No validation/test rows or additional synthetic dialogues were added.
• Canonical target fingerprint: 0a213de3804988cca9ade4b6a7235e981df1dd50a9cde5af6f0083c7d0bd87b8.
• Contextualized-data fingerprint: fd8c8c091f4bc0027c91b22c257cba30a2fd6c21004342f63c6029e82d83e934.

The ten prompt strategies were sampled deterministically with seed 11. Counts were: application 477, broader context 486, comparison 497, creative interpretation 478, critical analysis 500, depth 519, key concepts 483, questions 479, reflection 473, and synthesis 485.

Training configuration

The run used PyTorch 2.8.0+cu128, Transformers 5.2.0, and PEFT 0.19.1. It completed 610/610 optimizer steps without resume in 13,046 seconds. Final training loss was 1.3372108811. The training workload was patientagent-jackn-g43-cft-r16-train-173834 from PatientAgent commit d021583e4989327a023fd438235beab16c3bf5e9.

Validation and evaluation status

The saved adapter passed configuration, data-identity, completion, secret, full-model-file, and tensor-integrity checks. All 400 adapter tensors were finite. The adapter weights contain 30,474,240 parameters, are 121,951,008 bytes, and have SHA-256 dab128c9e366ca6ee17e5e6a124682a59c72bda2aad1e593fda974ad449eddb9.

Generation-only validation completed on the same source-eligible IDs used by the current PatientAgent comparison:

• Test1: 174 dialogues, 719 patient turns, cache fingerprint f7cb67e94239e78965fd4ba39df533d50ea166c66ecea6754dadd3dad6379ebd.
• Test2: 178 dialogues, 826 patient turns, cache fingerprint 8dc0b91946aa51c813baa260eb0f2e56a05ea9adc938cec306964270b5b0b4dc.

The caches passed exact ID/order, facts, turn-structure, unchanged-clinician-text, nonempty-patient-response, and prompt-leakage checks. Manual review found generally concise and context-sensitive outputs, with no systematic prompt leakage or refusal behavior. Some details in long conversations may still be unsupported and should be assessed by the matched evaluator.

Matched G-Eval is complete on the exact same source-eligible IDs as the current comparison. Scores use one Kimi-K2.7-Code judge pass per dialogue with a 100,000-token output ceiling. Hallucination and irrelevance are lower-is-better; anthropomorphism is higher-is-better.

All 352 scores completed without retry, split checkpoints matched the validated cache fingerprints, and an independent validator recomputed every split and combined mean. Against the canonical SFT16 row, the automated point estimates are slightly better for irrelevance (1.22 vs. 1.36) and anthropomorphism (8.67 vs. 8.56), but slightly worse for hallucination (0.72 vs. 0.70). This single run therefore does not establish uniform superiority or statistical significance.

Manual review of 12 representative or extreme CFT dialogues found 11 judge explanations broadly aligned with the rubric and one clear patient/non-patient role-attribution error in a mixed child/parent dialogue. No score was corrected after inspection. Treat these as noisy automated benchmark measurements, not clinical-quality evidence. The complete split conversations/results and combined validation bundle are pinned at patientagent-eval-results@cf538c3a3e9bed5e5f251405e9fc7b588ebc4664.

The matched source-versus-generated discriminator additionally covered the same frozen 53-dialogue/280-turn manifest used for every established model, with both A/B orders and analytical 0.5 scores for 21 text-identical pairs. CFT detection was 0.341 [0.294, 0.412] by patient-turn micro-average and 0.397 [0.321, 0.480] by dialogue macro-average, for a 0.659 fooling rate. All 518 API calls completed on the first attempt and passed independent recomputation. The 6.4-point detection increase relative to SFT16 was not resolved by the paired 95% cluster interval [-0.7, 11.9 points]. CFT responses were shorter than the references on average (30.8 versus 45.8 characters), and the judge was 21.4 points more accurate when the CFT response appeared in slot B; this metric therefore reflects source style and order effects rather than clinical quality. The complete merged discriminator bundle is pinned at patientagent-eval-results@cf96ac74aae731f737211c2da69ae0a810401ad1.

Loading the adapter

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "Qwen/Qwen3.5-4B"
adapter_id = "cs552-the-expendables/patientagent-cft-augmented-sft-r16"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base = AutoModelForCausalLM.from_pretrained(
    base_id,
    revision="851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
    device_map="auto",
    dtype="auto",
)
model = PeftModel.from_pretrained(base, adapter_id)
model.eval()

Use PatientAgent's normal case-facts and dialogue-history inference prompt. Do not prepend the contextual learning guidance at inference.

Limitations

• This is a single experimental run, not a hyperparameter sweep or an exact reproduction of the paper.
• The training corpus is small and English-only. Source and extracted-fact errors can propagate to the adapter.
• A patient simulator can invent, omit, or contradict clinical information. Outputs require independent evaluation and must never be used for clinical decision-making.
• Training loss is not a measure of clinical fidelity or conversation quality.
• The original CFT paper primarily studied high-information medical and financial corpora; effectiveness on this conditional dialogue adaptation was not established by the paper itself.
• G-Eval uses one LLM judge family and no human clinical adjudication; the manual audit found one clear role-attribution error among 12 inspected dialogues.
• The source discriminator has a material length/style and presentation-order confound; it measures detectability against MTS-Dialog wording, not human-likeness or clinical correctness.

Citation

@inproceedings{choi2025contextual,
  title={Teaching LLMs How to Learn with Contextual Fine-Tuning},
  author={Choi, Younwoo and Asif, Muhammad Adil and Han, Ziwen and Willes, John and Krishnan, Rahul G.},
  booktitle={International Conference on Learning Representations},
  year={2025},
  url={https://arxiv.org/abs/2503.09032}
}