Gemma-4-IT-SFT-RLVR-Medical

Gemma-4-it fine-tuned on PubMedQA, using SFT and RLVR.

Sample Output with Reasoning:

Question: Do GEC produce and bear factor H under complement attack?
Answer:

This study was the first to demonstrate that factor H was expressed as a protein on the membrane of GEC. Since GEC are an important target in the development of complement-mediated kidney injury, and factor H is involved in the ability of the kidney to produce factor H, the study thus provides a new perspective on the interactions between factor H and GEC.
Final Decision: yes

View Internal Thinking Process and Detailed Reasoning

1.  **Analyze the Request:** The user has provided a scientific context (a research abstract) and implicitly expects an answer (yes/no/maybe) based on the provided text.

2.  **Analyze the Context (Abstract):**
    *   **Topic:** Factor H in rat glomerular epithelial cells (GEC).
    *   **Methodology:** Used polyclonal antibody, PCR-based cloning, RT-PCR, immunofluorescence microscopy, Western blot, and Heymann nephritis model.
    *   **Key Findings:**
        *   Factor H was identified on cultured GEC membranes.
        *   Cloning GEC factor H cDNA showed it was full-length.
        *   Factor H mRNA increased under complement activation in cultured GEC and in the PHN model.

3.  **Determine the Core Claim/Conclusion:** The GEC produce and express factor H.

4.  **Relate Findings to the Question:** The question asks if GEC produce and bear factor H under complement attack. The findings directly support this conclusion.

5.  **Formulate the Reasoning:** Summarize the key experimental evidence that led to the conclusion. (GEC are a source of factor H.)

6.  **Final Check:** Does the answer address the context accurately and directly? Yes. (The study confirms a new role for GEC in factor H production).Despite its importance in the regulation of complement activation, the main source of factor H in the kidney is still unclear. Our study demonstrates that GEC produce factor H and the synthesized factor H is presented on the surface of these cells. We therefore conclude that GEC may be a source of factor H that contributes to the kidney's ability to resist complement attack. 
Final Decision: yes

Setup

Inference

Full Precision

#!pip install transformers, torch, accelerate
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("lukasdrews/Gemma-4-E2B-IT-SFT-RLVR-Medical")
model = AutoModelForCausalLM.from_pretrained("lukasdrews/Gemma-4-E2B-IT-SFT-RLVR-Medical")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "Do GEC produce and bear factor H under complement attack?"}
        ]
    },
]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=1024)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Quantized

# !pip install llama-cpp-python
from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="lukasdrews/Gemma-4-E2B-IT-SFT-RLVR-Medical-GGUF",
    filename="gemma-4-E2B-it-sft-rlvr-medical-Q4_K_M.gguf",
    verbose=False,
)
messages = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "Do GEC produce and bear factor H under complement attack?"}
        ]
    },
]

outputs = llm.create_chat_completion(messages, max_tokens=1024)
print(outputs["choices"][0]["message"]["content"])

Fine-tuning

Requirements:

• uv
• CMake (for quantization)

uv sync --no-dev &&
uv run QLORA_SFT.py && 
uv run GRPO_RLVR.py && 
uv run evaluate.py &&
./quantize_models.sh

Benchmarks

Gemma-4-E2B-it

Model	Quantization	PubMedQA (In-Domain)	MedQA-USMLE (Zero-Shot Transfer)
Gemma-4-E2B-it (base model)	-	58.10 %	29.54 %
Gemma-4-E2B-it + SFT + RLVR	-	73.10 %	43.05 %
Gemma-4-E2B-it + SFT + RLVR	Q8_0	72.40 %	43.00 %
Gemma-4-E2B-it + SFT + RLVR	Q6_K	72.10 %	42.18 %
Gemma-4-E2B-it + SFT + RLVR	Q5_K_M	72.00 %	38.88 %
Gemma-4-E2B-it + SFT + RLVR	Q4_K_M	71.80 %	38.88 %

Roadmap

• SFT + RLVR for Gemma-4-E2B
• GGUF Quantization for E2B
• Complete MedQA-USMLE Benchmarks for E2B (In Progress)
• Release E4B model and benchmarks