Maxx — qwen2.5-1.5b-agentic-v1
A fine-tuned version of Qwen2.5-1.5B-Instruct optimised for agentic tool-use and step-by-step reasoning.
Training
- Base model: Qwen2.5-1.5B-Instruct
- Method: QLoRA SFT → DPO (Unsloth)
- SFT data: OpenHermes 2.5, SlimOrca, Glaive FC v2, Hermes FC, synthetic ReAct tool-use trajectories (~131K examples)
- DPO data: Synthetic preference pairs — good vs bad tool calls (~5K pairs)
- Compute: Kaggle T4
Targets
- HuggingFace Open LLM Leaderboard
- Berkeley BFCL (Function Calling Leaderboard)
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
"bolajiev/maxx1.5Bv2",
torch_dtype="auto",
device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("bolajiev/qwen2.5-1.5b-agentic-v1")
- Downloads last month
- -
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support