A Qwen3.5-4B fine-tune that turns messy, free-form HR notes into clean, structured JSON — in Turkish or English, matching whatever language the input is written in.

Try the live demo →

What it does

HR teams take raw notes during disciplinary meetings, ethics hotline calls, internal audits, and performance reviews. This model structures those notes into the exact JSON fields an HR system needs, without inventing details that weren't actually stated.

Four supported document types:

Type	Turkish name	Example use case
Disciplinary record	Disiplin tutanağı	Structuring a workplace incident interview
Ethics hotline report	Etik hattı bildirimi	Anonymized whistleblower report intake
Internal audit finding	İç denetim bulgusu	Turning an auditor's raw note into a findings record
Performance consistency check	Performans tutarlılık kontrolü	Flagging inconsistencies in review notes

Key properties

Bilingual output matching: writes Turkish JSON for Turkish input, English JSON for English input.
Hallucination-averse: trained to state only what the note actually says, and to flag anything unverified rather than assert it as fact.
Privacy-conscious by default: prefers job title/role over personal name for unverified third parties (witnesses, observers) unless the schema specifically requires identifying the accused/reported party.
Small and CPU-deployable: 4B parameters, ships as a Q4_K_M GGUF that runs comfortably on CPU-only hardware (see the live demo, which runs on free CPU infrastructure).

Usage

llama.cpp / llama-cpp-python (recommended — this is what the live demo uses)

python

from llama_cpp import Llama
from transformers import AutoTokenizer

REPO_ID = "mustafabasar/ik-belge-asistani-qwen3.5-4b"

llm = Llama.from_pretrained(
    repo_id=REPO_ID,
    filename="ik-belge-asistani-qwen3.5-4b.Q4_K_M.gguf",
    n_ctx=4096,
    n_gpu_layers=0,  # CPU-only; raise this if you have a GPU
)
tokenizer = AutoTokenizer.from_pretrained(REPO_ID)

messages = [
    {"role": "system", "content": "You are an HR document assistant. Structure raw notes into only the requested JSON fields, without adding invented information. Never assert something you are not sure about."},
    {"role": "user", "content": "Structure this internal audit finding into the requested JSON fields; state only verified facts and flag anything unconfirmed.\n\nWarehouse B, Izmir. During the March inventory count, 47 units of SKU-2291 were found missing from the system record."},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)
out = llm(prompt, max_tokens=400, temperature=0.1, stop=["<|im_end|>"])
print(out["choices"][0]["text"])

transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer

REPO_ID = "mustafabasar/ik-belge-asistani-qwen3.5-4b"
model = AutoModelForCausalLM.from_pretrained(REPO_ID)
tokenizer = AutoTokenizer.from_pretrained(REPO_ID)

Important: always pass enable_thinking=False in apply_chat_template. This model was fine-tuned to answer directly with structured JSON; with thinking enabled it will emit reasoning text instead.

Training

Base model: unsloth/Qwen3.5-4B
Method: LoRA fine-tuning via Unsloth, merged to full weights
Precision: trained in float32 on a Kaggle T4 (Qwen3.5's architecture rejects float16 training; bf16 isn't supported on Turing-generation GPUs)
Quantization: merged fp16 weights + a Q4_K_M GGUF for CPU inference

Limitations

JSON field names stay in Turkish even when the input and values are in English (e.g. bulgu, kanit). These are fixed schema keys the model learned as structural constants rather than translatable text, not a bug — a fully bilingual schema would require retraining with an English-schema dataset.
Context window: 4096 tokens total, with 400 reserved for the model's output. Very long raw notes (roughly 1,800+ Turkish words or 2,400+ English words) may not fit; a friendly error message for this case is planned but not yet implemented in the demo.
Multi-token prediction / speculative decoding is not supported for this checkpoint (the bundled GGUF converter used for export does not yet map the Qwen3.5 MTP layer to GGUF).

License

Apache 2.0, inherited from the base model.

This qwen3_5 model was trained 2x faster with Unsloth and Hugging Face's TRL library.