Stage-1 checkpoint. Only the 2-layer MLP projector has been trained. The model can emit plain BLIP-style captions, but has not been instruction-tuned and is not intended as a deployable chat model — it is the pretraining foundation for downstream stage-2 fine-tunes.

For the stage-2 instruction-tuned bundle (same architecture, trained on LLaVA-Instruct-150K), see alpharomercoma/vqwen3-4b.

Loads as a stock LlavaForConditionalGeneration — no trust_remote_code.

Architecture

Vision tower: openai/clip-vit-large-patch14-336 — frozen
Projector: 2-layer MLP2xGELU (Linear 1024→2560 → GELU → Linear 2560→2560) — trained (only delta)
LLM: Qwen/Qwen3-4B — frozen

Trainable parameter count: ~9 M (two Linear layers). Everything else is loaded unchanged from its base checkpoint.

Training recipe

Dataset: liuhaotian/LLaVA-Pretrain — 558 K BLIP-captioned image–text pairs from LAION/CC/SBU. Conversation format is plain (no chat template, no system prompt): <image> on the human turn, caption on the assistant turn, loss masked on the human side.

Hyperparameters (single NVIDIA H200, bf16, SDPA, Liger-Kernel):


Global batch size	256
Learning rate	1e-3, cosine, warmup ratio 0.03
Weight decay	0.0
Epochs	1
Max sequence length	2048
Precision	bf16

What this model can do

Emit short captions for images when prompted in the training-time "plain" format.
Serve as a starting point for stage-2 instruction tuning (LoRA on the LLM
- continued projector training) — the frozen CLIP features are already aligned to Qwen3's embedding space.

What this model is NOT

Not an instruction-following model. Has never seen chat-formatted supervision, so it will not reliably follow yes/no-style prompts.
Not specialized for any domain — this is the generic alignment checkpoint, pre-any task fine-tuning.

Quick start

python

import torch
from PIL import Image
from transformers import LlavaForConditionalGeneration, AutoProcessor

model_id = "alpharomercoma/vqwen3-4b-pretrain"
model = LlavaForConditionalGeneration.from_pretrained(
    model_id, dtype=torch.bfloat16, device_map="auto"
).eval()
processor = AutoProcessor.from_pretrained(model_id)

image = Image.open("my_image.jpg").convert("RGB")
messages = [{
    "role": "user",
    "content": [
        {"type": "image"},
        {"type": "text", "text": "Describe this image."},
    ],
}]
prompt = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = processor(text=prompt, images=image, return_tensors="pt").to(model.device)
inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16)
out = model.generate(**inputs, max_new_tokens=64, do_sample=False)
print(processor.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Pairs with

Stage	Raw delta (Kaggle)	Bundled (HF)
Stage-1 alignment	`vqwen-projector`	this model
Stage-2 instruction	`vqwen-lora`	`vqwen3-4b`

Credits

Base vision: openai/clip-vit-large-patch14-336
Base LLM: Qwen/Qwen3-4B
Dataset: liuhaotian/LLaVA-Pretrain
Recipe: LLaVA-1.5 stage-1 alignment

License

Apache 2.0 for the trained projector. Base models retain original licenses.