This repository contains a full fine-tuned model (not LoRA adapter) based on Qwen3-4B-Instruct-2507, trained with multi-turn agentic SFT using the Open-AgentRL framework (verl FSDP SFT Trainer).

Note: This is the final checkpoint (epoch 10 of 10).

Training Objective

This model is trained to acquire multi-turn agent reasoning with tool use — specifically, the ability to iteratively call a code_interpreter tool to solve math and coding problems.

Loss is applied to all assistant turns in the multi-turn trajectory, enabling the model to learn the full agentic loop: Think → Code → Execute → Observe → Answer.

This SFT stage serves as a cold-start for subsequent GRPO reinforcement learning.

Training Configuration

Parameter	Value
Base model	`Qwen/Qwen3-4B-Instruct-2507`
Method	Full fine-tuning (FSDP, bfloat16)
Max sequence length	32,768
Epochs	10
Train batch size	16
Micro batch size per GPU	1
Truncation	right
Trainer	`verl.trainer.fsdp_sft_trainer`

Dataset

Name: y-ohtani/open_agentrl_like_sft
License: Apache-2.0
Samples: 2,000 multi-turn conversations
Source: Derived from swordfaith/ReTool-SFT-multi-turn (Apache-2.0)
Domain: Mathematical reasoning with code interpreter
Structure: Average 7.0 messages per conversation (user: 2K, assistant: ~7K, tool: ~5K)

All training data is sourced from Apache-2.0 licensed datasets. This repository does NOT redistribute the dataset.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "y-ohtani/qwen3-4b-ra-sft-merged"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "user", "content": "Solve the equation x^2 - 5x + 6 = 0 step by step."}
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=2048)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Sources & Terms

Component	Source	License
Base model	Qwen/Qwen3-4B-Instruct-2507	Apache-2.0
SFT dataset	y-ohtani/open_agentrl_like_sft	Apache-2.0
Training framework	Open-AgentRL (verl)	Apache-2.0

Users must comply with the base model license and dataset terms.

Intended Use & Limitations

Intended: Agentic reasoning tasks with tool use (math, coding). This model is designed as an intermediate checkpoint for further RL training (GRPO-TCR).
Not intended: Production deployment without further evaluation.
Limitations: Performance on non-math/non-coding tasks may degrade compared to the base instruct model.