QLoRA adapter (r=16) on top of Qwen2.5-Coder-7B-Instruct, fine-tuned on curated high-quality reasoning traces. Focus: explaining the reasoning before writing code, honesty about uncertainty (anti-hallucination), and Indonesian developer context. All model output is in Indonesian.

Trained capabilities

Chain-of-thought: root cause → alternatives + trade-offs → solution
Self-correction: draft → mental test → find its own bug → fix
Multi-turn debugging, asking clarification on ambiguous requests, refusing misguided requests
Web-search behavior: emits [CARI: <query>], then answers from an injected [HASIL PENCARIAN] block

Usage

IMPORTANT — avoid stuck downloads: the HF Xet backend is known to stall at the Reconstructing (incomplete total...) stage (xet-core#850). Disable Xet before any imports:

python

import os
os.environ["HF_HUB_DISABLE_XET"] = "1"  # MUST be the very first cell/line

Via Unsloth (recommended — runs on free Colab T4)

bash

pip uninstall -y hf-xet
pip install unsloth

python

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "TheCoderScientist/GarudaCoder-7B-Coder-ID-lora",
    max_seq_length=4096, load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

pesan = [
    {"role": "system", "content": "Kamu adalah GarudaCoder, asisten coding berbahasa Indonesia yang teliti, menjelaskan alasan sebelum menulis kode, dan tidak pernah mengarang fakta."},
    {"role": "user", "content": "Kenapa [[0]*3]*3 di Python bikin semua baris ikut berubah?"},
]
inputs = tokenizer.apply_chat_template(pesan, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")
out = model.generate(input_ids=inputs, max_new_tokens=1024, temperature=0.3, do_sample=True)
print(tokenizer.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))

Via transformers + PEFT (without Unsloth)

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel

bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4",
                         bnb_4bit_compute_dtype=torch.float16)
base = AutoModelForCausalLM.from_pretrained(
    "unsloth/Qwen2.5-Coder-7B-Instruct", quantization_config=bnb, device_map="auto")
model = PeftModel.from_pretrained(base, "TheCoderScientist/GarudaCoder-7B-Coder-ID-lora")
tokenizer = AutoTokenizer.from_pretrained("unsloth/Qwen2.5-Coder-7B-Instruct")

Important notes

Use the system prompt above verbatim — the model was trained with it; without it, reasoning quality degrades.
Sampling used during eval: temperature=0.3, do_sample=True, max_new_tokens=1024–1536.
VRAM: ~6 GB in 4-bit → fits on a T4 / free Colab. Do not load full fp16 on a T4 (needs ~16 GB).
This is a LoRA adapter (PEFT), not a full model — model.safetensors in this repo is small; the Qwen2.5-Coder-7B base model is downloaded automatically on load.
[CARI: ...] behavior: when the model needs up-to-date information it writes that tag, then answers from a [HASIL PENCARIAN] ... block you inject into the next turn. Without a search harness, just ignore the tag.

Troubleshooting

Symptom	Cause	Fix
Download stuck at `Reconstructing (incomplete total...)`	hf_xet stall bug	`HF_HUB_DISABLE_XET=1` before imports, or `pip uninstall hf-xet`
`CUDA out of memory` on T4	loaded without 4-bit	make sure `load_in_4bit=True`
English / shallow answers	system prompt not used	use the system prompt above verbatim
`KeyError` / adapter won't attach in PEFT	wrong base model	use `unsloth/Qwen2.5-Coder-7B-Instruct` (or `Qwen/Qwen2.5-Coder-7B-Instruct`)

Training

QLoRA 4-bit, r=16 alpha=16, 7 projection layers, max_seq 4096, T4 (fp16)
SFT with train_on_responses_only + DPO (chosen=reasoning/honest, rejected=code-dump/hallucination)
On-policy loop: the model is evaluated on 40 questions → graded on a 5-criteria rubric → gold corrections are written for each error → becomes next round's training data
Dataset: TheCoderScientist/garudacoder-id-coding