kinkinkin123456/nexus-qwen35-0p8b-smoke-merged

Nexus Qwen3.5-0.8B Smoke Merged Model

This repository contains a standalone merged checkpoint produced from the Nexus Qwen3.5-0.8B smoke LoRA adapter.

It is intended for teammates who want a full model directory instead of loading a separate PEFT adapter.

What This Repo Contains

This is a merged full-weight checkpoint.

Main files:

• —model.safetensors
• —config.json
• —processor_config.json
• —tokenizer.json
• —tokenizer_config.json
• —chat_template.jinja
• —nexus_configs/

Training Summary

• —Base model: Qwen/Qwen3.5-0.8B
• —Source adapter: Nexus Stage 1 smoke LoRA finetune artifact
• —Merge type: PEFT adapter merged into base model
• —Training goal: smoke validation for training, save, reload, and inference paths
• —Data: a small HatefulMemes smoke subset
• —Steps: 10
• —Precision: bf16

Recommended Usage

Inside the Nexus codebase, point both model and processor to this directory:

from Nexus import MultimodalEmbedder

model = MultimodalEmbedder(
    model_name_or_path="path/to/this/merged/repo",
    processor_name_or_path="path/to/this/merged/repo",
    model_type="qwen3_5",
    trust_remote_code=True,
    normalize_embeddings=True,
    pooling_method="last_token",
)

Limitations

• —This is still a smoke-stage artifact.
• —It is useful for code verification and release testing, not for final capability claims.
• —The underlying finetune is intentionally tiny.