The model is trained to follow the ReAct paradigm: Thought → Action → Observation → Final Answer. After training, the model stops after outputting an Action and waits for real tool results, rather than hallucinating fake observations. When tools return errors, it can switch strategies.

Training Details

Parameter	Value
Base Model	Qwen2.5-7B-Instruct
Method	QLoRA (4-bit, R=32, Alpha=64)
Training Samples	8,640 (10% negative/error recovery)
Epochs	2.2 (best eval_loss=0.1766 at step 1,000)
GPU	RTX 4070 SUPER (12GB, 9.8GB used)
Framework	Unsloth + Transformers + PEFT

Evaluation Results

Model	Success Rate	Obs. Hallucination
FT_ReAct (Ours)	92.0%	0%
Baseline_B (Instruct direct)	88.0%	0%
Baseline_A (Instruct + ReAct + FS)	83.0%	1%

Task Type	Baseline_B	Baseline_A	FT_ReAct
Single-tool (35)	94%	91%	97%
Multi-tool (35)	91%	91%	97%
Error Recovery (30)	77%	63%	80%

Usage

python

from unsloth import FastLanguageModel
from peft import PeftModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "unsloth/Qwen2.5-7B-Instruct-bnb-4bit",
    max_seq_length=2048, dtype=None, load_in_4bit=True,
)
model = PeftModel.from_pretrained(
    model, "Yuh4ng/react-qwen2.5-7b-qlora-4070s"
)
FastLanguageModel.for_inference(model)

Limitations

Trained on mock tools, not real APIs
5 tool types covered (weather, stock, search, calculate, time)
12GB VRAM limits context to ~1,536 tokens
May occasionally hallucinate numbers in final answer

License

MIT