Instructions to use idalabs/IDA-4B-v1.0.2-P4.9 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use idalabs/IDA-4B-v1.0.2-P4.9 with PEFT:
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B") model = PeftModel.from_pretrained(base_model, "idalabs/IDA-4B-v1.0.2-P4.9") - Notebooks
- Google Colab
- Kaggle
IDA 4B v1.0.2 β P4.9 Β· LoRA Adapter
Indonesian-focused assistant fine-tuned by IDA Labs from Qwen/Qwen3-4B. This repository contains the P4.9 LoRA adapter; it is not a standalone full model. The merged quantized version is available separately as IDA 4B v1.0.2 P4.9 β GGUF Q4_K_M.
Model at a glance
| Item | Value |
|---|---|
| Adapter format | PEFT LoRA (adapter_model.safetensors) |
| Base required | Qwen/Qwen3-4B |
| Training method | QLoRA, 4-bit NF4 |
| Training | 2 epochs, 138 optimizer steps |
| Private training / validation | 548 / 112 examples |
| Final validation loss | 1.5901 |
| GGUF for local use | Download standalone Q4_K_M |
Load with Transformers + PEFT
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
base_id = "Qwen/Qwen3-4B"
adapter_id = "idalabs/IDA-4B-v1.0.2-P4.9"
tokenizer = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto", torch_dtype="auto")
model = PeftModel.from_pretrained(base, adapter_id)
messages = [{"role": "user", "content": "Perkenalkan diri secara singkat."}]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, enable_thinking=False, return_tensors="pt").to(model.device)
output = model.generate(inputs, max_new_tokens=128, do_sample=False)
print(tokenizer.decode(output[0][inputs.shape[-1]:], skip_special_tokens=True))
Evaluation dashboard
The 24-prompt fresh private blind test was scored by an AI assistant, not an independent human panel. The 20/24 result applies to the P4.9 adapter before GGUF quantization, not to an independently rescored GGUF. Unseen blind prompts were excluded from training.
| Category | Pass / total | Pass rate |
|---|---|---|
| Identity & attribution | 4 / 4 | 100% |
| Instruction following | 6 / 7 | 85.7% |
| Facts & arithmetic | 5 / 6 | 83.3% |
| Drafting | 2 / 4 | 50.0% |
| Safety & action honesty | 3 / 3 | 100% |
| Overall | 20 / 24 | 83.3% |
Known issues: an incorrect sky-color answer, an incorrect odd-number classification, an invented claim that IDA Labs provides outsourcing services, and unsolicited impersonation of project Principal Ary Wibowo in a drafted presentation. Drafting missed its 60% threshold; high-risk false claims failed the release-quality gate. The project owner chose to publish this version as an experimental snapshot, with correction planned for a later version.
EOS completion (technical inference check)
EOS comparison covers completion of the 24 inference responses for each arm (base 22/24, P4.9 adapter 24/24). It does not establish a semantic quality improvement over the base model. GGUF had a separate successful llama.cpp CPU inference smoke test.
Training and data overview
| Training detail | Recorded value |
|---|---|
| Base model | Qwen/Qwen3-4B |
| Fine-tuning | QLoRA, 4-bit NF4 base during training |
| Epochs | 2 |
| Optimizer steps | 138 |
| Private training examples | 548 |
| Private validation examples | 112 |
| Final validation loss | 1.5901 |
The loss curves come from 27 logged training-loss points and 7 validation-loss points in the P4.9 Kaggle training logs. They are not external benchmark measurements. Dataset contents and blind-test questions remain private; these charts share only aggregate data. Supporting data: training-loss-aggregates.csv and evaluation-aggregates.json.
Limitations and responsible use
- Experimental release, not production-certified. Review critical facts, identity claims, and drafted business content before use.
- The 83.3% pass rate is from a small, assistant-reviewed private blind set, not independent human review or a recognized public benchmark.
- This release has known simple factual errors and attribution/persona mistakes; quantization does not correct them.
- Training and evaluation source prompts are private. No personal records or private raw dataset are included in this repository.
Credits
Fine-tuning and development: IDA Labs
Project Principal: Ary Wibowo
Base model: Qwen3-4B, Qwen Team / Alibaba Cloud. IDA Labs does not claim authorship of the Qwen base weights.
- Downloads last month
- 36



