qwen2.5-7b-agent-trajectory-mixed_dbv4_alfv4_1to1

This repository provides a merged full model fine-tuned for AgentBench tasks
(ALFWorld + DBBench).

Base model: Qwen/Qwen2.5-7B-Instruct

This repository contains fully merged model weights (LoRA merged into the base model).

Training Objective

This model is optimized for:

• Sequential trajectory planning (ALFWorld)
• Structured reasoning and database querying (DBBench)
• Deterministic action generation
• Reduced invalid action rate

Datasets Used

The model was trained using only officially provided training datasets:

• u-10bei/sft_alfworld_trajectory_dataset_v5
• u-10bei/dbbench_sft_dataset_react_v4

Mixing strategy:

• ALFWorld (v5) and DBBench (v4) mixed in a 1:1 ratio.
• No validation or test splits were used for training.

Fine-tuning Method

• Supervised Fine-Tuning (SFT)
• LoRA-based training
• LoRA weights merged into base model before upload
• Loss applied only to assistant outputs
• No external datasets were used

Reproducibility

Base model: Qwen/Qwen2.5-7B-Instruct

Training framework:

• Hugging Face Transformers
• PEFT (LoRA)

Evaluation decoding configuration:

• do_sample=False
• temperature=0.0
• Deterministic generation

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "HamadaMayu/qwen2.5-7b-agent-trajectory-mixed_dbv4_alfv4_1to1"

tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained( model_id, device_map="auto", )

prompt = "Your task prompt here" inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

output = model.generate( **inputs, max_new_tokens=512, do_sample=False, temperature=0.0, )

print(tokenizer.decode(output[0], skip_special_tokens=True))

Intended Use

• AgentBench evaluation
• Research on trajectory learning
• Educational experiments

Limitations

• Performance may degrade outside AgentBench domains.
• Long-horizon planning is limited by context length.
• Invalid actions may still occur under distribution shift.