---
base_model: Qwen/Qwen2.5-7B-Instruct
datasets:
- u-10bei/sft_alfworld_trajectory_dataset_v5
- u-10bei/dbbench_sft_dataset_react_v4
language:
- en
license: apache-2.0
pipeline_tag: text-generation
tags:
- lora
- merged
- agent
- tool-use
- alfworld
- dbbench
---

# Qwen2.5-7B-Instruct-SDFT-2ep-fp16

This repository provides a fine-tuned model based on **Qwen/Qwen2.5-7B-Instruct**. 
The model was initially trained using **LoRA + Unsloth** and has been **merged with the base model**. 
The weights in this repository are saved in **fp16** format, so you can load and use it directly without needing to load the base model and adapter separately.

## Training Objective

This model is trained to improve **multi-turn agent task performance**
on ALFWorld (household tasks) and DBBench (database operations).

Loss is applied to **all assistant turns** in the multi-turn trajectory,
enabling the model to learn environment observation, action selection,
tool use, and recovery from errors.

## Training Configuration

- Base model: Qwen/Qwen2.5-7B-Instruct
- Method: LoRA (merged into base model)
- Precision: fp16
- **Experimental Methods:** SDFT & Epiplexity *(Note: Implementation is still a work in progress)*
- Max sequence length: 4096
- Epochs: 2
- Learning rate: 2e-06
- LoRA: r=64, alpha=128

## Experimental Features

This version incorporates experimental training techniques, specifically **SDFT** and **Epiplexity**. 
However, the integration of these methods is not yet fully completed. We are still evaluating their impact on the model's reasoning capabilities and plan to refine them in future updates.

## Usage

You can load this model directly using `AutoModelForCausalLM`.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "aolans/Qwen2.5-7B-Instruct-SDFT-2ep-fp16"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
)
```

## References

The experimental training methods (SDFT and Epiplexity) applied in this model are based on the following research:

* [Self-Distillation Enables Continual Learning](https://arxiv.org/abs/2601.19897)
* [From Entropy to Epiplexity: Rethinking Information for Computationally Bounded Intelligence](https://arxiv.org/abs/2601.03220)

## Sources & Terms (IMPORTANT)

Training data: u-10bei/sft_alfworld_trajectory_dataset_v5, u-10bei/dbbench_sft_dataset_react_v4

Dataset License: MIT License. This dataset is used and distributed under the terms of the MIT License.
Compliance: Users must comply with the MIT license (including copyright notice) and the base model's original terms of use.