corbt's picture
Upload folder using huggingface_hub
00f1d2a verified
metadata
license: apache-2.0
base_model: unsloth/Qwen3-30B-A3B-Instruct-2507
tags:
  - qwen3
  - chat
  - instruct

OpenPipe/Qwen3-30B-A3B-Instruct-2507

This is a copy of unsloth/Qwen3-30B-A3B-Instruct-2507 with a fixed chat template for SFT training.

Changes from the original

The only change is the chat template. The original Qwen3 chat template adds <think></think> tags inconsistently - only to the last assistant message in a conversation, not to earlier assistant messages in the history.

This causes issues during SFT training because:

  • Historical assistant messages: <|im_start|>assistant\nHi there!<|im_end|>
  • Target assistant message: <|im_start|>assistant\n<think>\n\n</think>\n\nI am well!<|im_end|>

The model learns an inconsistent pattern, leading to degraded outputs.

Fixed template

This model uses a consistent chat template that adds <think>\n\n</think>\n\n to ALL assistant messages:

{{- '<|im_start|>assistant\n<think>\n\n</think>\n\n' + content }}

This ensures the model learns a single, consistent pattern during training.

Usage

Use this model as your base model for SFT training with Qwen3-30B-A3B when you need consistent chat formatting.

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("OpenPipe/Qwen3-30B-A3B-Instruct-2507")
model = AutoModelForCausalLM.from_pretrained("OpenPipe/Qwen3-30B-A3B-Instruct-2507")

Model Details