tbench-qwen-sft-v3-epoch3

tbench-qwen-sft-v3-epoch3

Supervised fine-tune of Qwen/Qwen3-8B for terminal-agent / shell-tool-use tasks.

Checkpoint at epoch 3 of the merged-v3 training run.

Training data

1,112 trajectories combining:

• Kimi-K2 thinking traces (v10) on terminal-bench tasks
• seta-env synthetic data (filtered)

Inference

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tokenizer = AutoTokenizer.from_pretrained("alievak/tbench-qwen-sft-v3-epoch3")
model = AutoModelForCausalLM.from_pretrained(
    "alievak/tbench-qwen-sft-v3-epoch3",
    dtype=torch.bfloat16,
    device_map="auto",
)

The model produces <think>...</think> blocks before tool calls. Use the included chat_template.jinja for proper rendering.

Notes

• Architecture: Qwen3ForCausalLM (same as base Qwen/Qwen3-8B, ~8B params)
• Precision: bfloat16
• Behavior on general chat / instruction following inherits from base Qwen3-8B