---
license: apache-2.0
datasets:
- Azure99/blossom-v6.3-sft-stage1
- Azure99/blossom-v6.3-sft-stage2
language:
- zh
- en
base_model:
- Qwen/Qwen3.5-35B-A3B-Base

---

# **BLOSSOM-V6.4-35B-A3B**

[💻Github](https://github.com/Azure99/BlossomLM) • [🚀Blossom Chat Demo](https://blossom-chat.com/)

### Introduction

Blossom is a powerful open-source conversational large language model that provides reproducible post-training data, dedicated to delivering an open, powerful, and cost-effective locally accessible general-purpose model for everyone.

The Blossom-V6.4 series largely follows the V6.3 training recipe and uses the same training data, with a small number of multimodal samples added to preserve the multimodal capabilities of the Base models.

You can find the training data here: [Blossom-V6.3-SFT-Stage1](https://huggingface.co/datasets/Azure99/blossom-v6.3-sft-stage1) (1 epoch)、[Blossom-V6.3-SFT-Stage2](https://huggingface.co/datasets/Azure99/blossom-v6.3-sft-stage2) (3 epoch).

### **Data Synthesis Workflow Overview**

Primarily employs three cost-effective models: Deepseek-V3.1, Gemini 2.5 Flash, and Qwen3-235B-A22B-Instruct-2507 (denoted as A, B, C)—to regenerate responses under different scenarios using tailored synthesis strategies.

For example:

- In objective scenarios like mathematics (where answers are unique), Model A first generates responses as a "teacher." If reference answers exist in the source data, Model B verifies the correctness of A's responses against them. If no reference answers exist, Model C generates a second response, and Model B checks consistency between A and C's outputs. Inconsistent responses are filtered out.
- For subjective scenarios, three models cross-evaluate each other. For instance, Models A and B generate responses to a question, and Model C evaluates which is better. The superior response may be retained as training data or used for preference data construction. To mitigate model bias, roles (respondent/evaluator) are randomly assigned to A, B, and C in each instance.

Additional rule-based filtering is applied, such as:

- N-Gram filtering to remove data with many repetitions.
- Discarding questions containing toxic content that triggers teacher model refusals.

Further technical details will be released in the future. The data is synthesized by the [🌸BlossomData](https://github.com/Azure99/BlossomData) framework.

### Inference

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL = "Azure99/Blossom-V6.4-35B-A3B"

model = AutoModelForCausalLM.from_pretrained(MODEL)
tokenizer = AutoTokenizer.from_pretrained(MODEL)

messages = [
    {"role": "user", "content": "北京有什么好吃的"}
]

inputs = tokenizer.apply_chat_template(
    messages,
    return_tensors="pt",
    return_dict=True,
).to(model.device)

generated_ids = model.generate(**inputs, max_new_tokens=512)
generated_ids = generated_ids[:, inputs["input_ids"].shape[-1]:]

print(tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0])
```