---
library_name: peft
base_model: Qwen/Qwen3-4B
tags:
- lora
- sft
- transformers
- trl
- function-calling
- telco
pipeline_tag: text-generation
---

# 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](https://huggingface.co/datasets/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](https://wandb.ai/tiendungcbh1801-hanoi-university-of-science-and-technology/telco-fc/runs/fhjp6hke)

## Output format

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