CT4nk3r/gemma4-e4b-deepresearcher-lora

Gemma4 E4B Researcher LoRA

This is an experimental LoRA adapter that specializes google/gemma-4-e4b-it for research-assistant behavior: evidence-aware answers, citation discipline, uncertainty when evidence is insufficient, and concise synthesis.

This adapter is not an official Google, Alibaba, Tongyi, or DeepResearch release.

Training summary

• Base model: google/gemma-4-e4b-it
• Method: supervised fine-tuning with PEFT/LoRA
• Teacher model: Alibaba-NLP Tongyi DeepResearch served locally through LM Studio
• Bootstrap data: public WebGPT-style cited QA examples converted into chat SFT format
• Distillation goal: rewrite examples into source-grounded researcher responses with explicit citations and insufficient-evidence behavior

Intended use

Use this adapter with the Gemma4 E4B base model for local research-assistant experiments where the application supplies sources, notes, or retrieved evidence and expects grounded synthesis.

Limitations

• The adapter does not add factual knowledge by itself.
• The model can still hallucinate, misuse citations, or overstate weak evidence.
• It should be paired with retrieval, source tracking, and citation validation.
• It is not suitable for high-stakes research without human review.

Loading

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "google/gemma-4-e4b-it"
adapter = "<your-username>/<repo-name>"

tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base)
model = PeftModel.from_pretrained(model, adapter)

License and data notes

Users must comply with the base model license and the terms of any datasets used during training. The training data may include public cited-QA examples and teacher-distilled outputs; verify redistribution rights before publishing data.

Progression snapshot

• Snapshot: progression-20260602-cumulative-relay
• Source GitHub release: https://github.com/CT4nk3r/gemma4-deepresearcher/releases/tag/progression-20260602-cumulative-relay
• Clean SFT examples: 830
• Raw teacher examples: 1032
• Adapter source: runs\gemma4-e4b-deepresearch-lora-latest after the first cumulative --resume-adapter relay pass on 2026-06-02.
• Note: training is still experimental and continuing locally; this is a progress checkpoint, not a final model.