Modelpublic

Jun1801/telco-fc-m1b-qwen3-4b

sourceHugging Faceupdated 3mo agoView on Hugging Face

0likes7downloads

Model Card

M1b — Telco Function Calling SFT (Qwen3-4B + LoRA)

LoRA adapter fine-tuned on 26 real Viettel KPI functions (Vietnamese, read-only).

Results (1806 eval samples)

Split	Accuracy
abstain	100.0%
missing_slot	99.4%
unseen	97.4%
masked	94.6%
seen	94.0%
parallel	80.2%
multi_step	54.3%
Overall	84.9%

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B", torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(base, "Jun1801/telco-fc-m1b-qwen3-4b")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B")

Training

—Base model: Qwen/Qwen3-4B
—Dataset: Jun1801/telco-fc-dataset — 7,526 samples (real KPI + warmup)
—Method: SFT with TRL SFTTrainer, LoRA (r=16, alpha=32)
—Epochs: 3, LR: 2e-4 (cosine decay)
—Final train loss: 0.088 | eval loss: 0.012
—W&B: run fhjp6hke

Output format

json

{"action": "call_function", "call": {"tool_name": "...", "arguments": {...}}}
{"action": "call_functions", "calls": [{...}, ...]}
{"action": "ask_clarification", "asked_slots": ["..."]}
{"action": "abstain", "reason": "..."}