adityag-india

qwen25-1.5b-medical-o1-qlora

Deploy Dedicated

README

License: apache-2.0

Base model

Base model: Qwen/Qwen2.5-1.5B-Instruct
Architecture: decoder-only Qwen2.5 causal language model
Adapter method: QLoRA / PEFT LoRA
Quantization during training: 4-bit NF4
Training framework: Unsloth + TRL + PEFT

The base model is distributed under Apache-2.0.

Dataset

Training used the English configuration of:

FreedomIntelligence/medical-o1-reasoning-SFT

Dataset fields used:

Question
Complex_CoT
Response

The complete dataset contained 19,704 examples and was split deterministically:

Table with columns: Split, Examples, Purpose
Split	Examples	Purpose
Train	18,704	Gradient updates
Validation	500	Checkpoint/generalization monitoring
Test	500	Final held-out comparison
Total	19,704

The dataset is distributed under Apache-2.0 by its publisher.

Training format

Each example was formatted using the Qwen chat template in the following logical form:

text

system:You are a medical reasoning assistant.user:<Question>assistant:<reasoning><Complex_CoT></reasoning><Response>

The <reasoning> tags were plain text delimiters in the training content.

Training configuration

Table with columns: Parameter, Value
Parameter	Value
Epochs run	5
Selected checkpoint	Epoch 2 / step 2118
Max sequence length	1536
Physical batch size	4
Gradient accumulation	2
Effective batch size	8
Learning rate	2e-4
LR scheduler	Cosine
Warmup steps	75

The LoRA adapter contained approximately 36.9M trainable parameters.

Validation behavior

The original 5-epoch run reached its minimum validation loss at Epoch 2:

Table with columns: Epoch, Validation loss
Epoch	Validation loss
1	1.437329
2	1.423186
3	1.443278
4	1.496967
5	1.549110

Validation performance improved through Epoch 2 and then degraded through Epochs 3-5.

Held-out test results

The untouched 500-example test split was evaluated using the same formatting and max_sequence_length=1536. No test examples were truncated.

Table with columns: Model, Test loss, Perplexity, Mean token accuracy
Model	Test loss	Perplexity	Mean token accuracy
Base Qwen2.5-1.5B-Instruct	2.006018	7.4337	0.551078
Medical adapter — Epoch 2	1.452841	4.2752	0.623556
Medical adapter — Epoch 5	1.582141	4.8654	0.613223

Relative to the base model, the Epoch 2 adapter reduced test loss by about 27.58% and increased mean token accuracy by about 7.25 percentage points.

Continuing from Epoch 2 to Epoch 5 increased test loss by about 8.90% and reduced mean token accuracy by about 1.03 percentage points.

Multi-seed replication

To test whether the Epoch 2 result was simply a lucky training trajectory, the same training configuration and the same fixed train/validation split were repeated with five training seeds:

42, 123, 456, 789, 2026

Epoch 2 was the best validation checkpoint in 5/5 runs.

Table with columns: Training seed, Best epoch, Best validation loss
Training seed	Best epoch	Best validation loss
42	2	1.423093
123	2	1.422025
456	2	1.422589
789	2	1.422546
2026	2	1.422715

Mean Epoch-2 validation loss across these runs was approximately 1.422594 ± 0.000384 SD.

This supports training-seed robustness for the observed Epoch-2 optimum under this fixed data split and hyperparameter configuration. It does not establish that Epoch 2 is universally optimal under different data splits, datasets, models, or hyperparameters.

Using the adapter

Install the required libraries:

bash

pip install -U transformers peft accelerate torch

Example:

python

import torchfrom transformers import AutoTokenizer, AutoModelForCausalLMfrom peft import PeftModel BASE_MODEL = "Qwen/Qwen2.5-1.5B-Instruct"ADAPTER = "adityag-india/qwen25-1.5b-medical-o1-qlora" tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL) base_model = AutoModelForCausalLM.from_pretrained(    BASE_MODEL,    torch_dtype="auto",    device_map="auto",) model = PeftModel.from_pretrained(base_model, ADAPTER)model.eval() messages = [    {        "role": "system",        "content": "You are a medical reasoning assistant."    },    {        "role": "user",        "content": "What are common causes of iron-deficiency anemia?"    }] inputs = tokenizer.apply_chat_template(    messages,    add_generation_prompt=True,    tokenize=True,    return_tensors="pt",).to(model.device) with torch.inference_mode():    outputs = model.generate(        inputs,        max_new_tokens=512,        do_sample=False,    ) print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Evaluation scope and limitations

The reported evaluation metrics are token-level language-model metrics. They measure likelihood/perplexity and next-token accuracy on the held-out Medical-O1 test split.

They do not directly establish:

clinical correctness,
diagnostic safety,
treatment quality,
hallucination rate,
calibration,
performance on real patient records,
performance across medical specialties,
robustness to adversarial or ambiguous prompts.

The dataset includes generated reasoning traces, so the adapter may learn patterns and stylistic properties of those traces in addition to medical content.

The model should therefore be treated as a research artifact and not as a clinical system.

Reproducibility notes

The main experimental artifacts retained with this repository are intentionally small:

selected PEFT adapter weights,
adapter configuration,
final held-out test comparison,
dataset split metadata,
training summary,
per-seed summary JSON files.

Large optimizer states, scheduler states, caches, intermediate checkpoints, downloaded base-model weights, raw dataset copies, and training logs are intentionally omitted from this repository.

Licenses and upstream resources

Base model: Qwen/Qwen2.5-1.5B-Instruct — Apache-2.0
Dataset: FreedomIntelligence/medical-o1-reasoning-SFT — Apache-2.0

Please review the upstream model and dataset repositories for their full terms, documentation, and attribution requirements.

Reference Research Papers

@inproceedings{bui2025seeds, title = {Assessing the Macro and Micro Effects of Random Seeds on Fine-Tuning Large Language Models}, author = {Bui, Nghia T. and Savova, Guergana and Wang, Lijing}, booktitle = {Proceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics}, pages = {41--46}, year = {2025}, url = {https://aclanthology.org/2025.ijcnlp-short.3.pdf} }

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.