qwen3-4b-agent-trajectory-merged2

qwen3-4b-agent-trajectory-merged1

This repository provides a MERGED (fully materialized) model created by merging a LoRA adapter into the base model:

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Method: LoRA + Unsloth, then merge_and_unload() into full weights

✅ This repo contains the merged model weights. You do NOT need PEFT / LoRA adapters at inference time.

Note: This is a derivative of the base model. Usage must comply with the base model's original terms.

What is included

• Full merged model weights (the LoRA adapter is already merged)
• Tokenizer / config files needed for inference

Training Objective

This model is trained to improve multi-turn agent task performance on agent-trajectory style data (e.g., ALFWorld household tasks and DBBench database operations, depending on the datasets listed below).

Loss is applied to all assistant turns in the multi-turn trajectories so the model learns: environment observation, action selection, tool use, and recovery from errors.

Training Configuration (summary)

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Method: LoRA training (base loaded without 4-bit quantization), then merged into full weights
• Max sequence length: 2048
• Epochs: 2
• Learning rate: 1e-05
• LoRA: r=64, alpha=128

Usage

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

repo = "your_id/your-merged-repo"

tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    repo,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True,
)

# (optional) generation example
inputs = tokenizer("Hello! What should I do next?", return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(out[0], skip_special_tokens=True))