MiniCPM-V-4.7-1B-Pashto-Base

This is the Stage 2 base checkpoint of the nassimjp/MiniCPM-V-4.7-1B-Pashto model, fine-tuned from openbmb/MiniCPM-V-4.7-1B.

In this stage, the 18 newly added Pashto-specific atomic tokens and their embedding/lm_head weights (which were initialized randomly in Stage 1) have been aligned and harmonized with the surrounding Arabic-script embedding space, preparing the model for downstream multimodal and text fine-tuning.

• Developed by: nassimjp
• License: apache-2.0
• Finetuned from model: nassimjp/MiniCPM-V-4.7-1B-Pashto
• Base Model: openbmb/MiniCPM-V-4.7-1B
• Architecture: MiniCPMV4_7ForConditionalGeneration

Training & Alignment Details

• Stage: Stage 2 Alignment (Embedding & Initial Adaptation)
• Framework: Unsloth / Hugging Face TRL
• Precision: 16-bit (FP16 / BF16)

This minicpmv4_7 model was trained 2x faster with Unsloth and Huggingface's TRL library.

---

About MiniCPM-V 4.7 (Base Architecture Overview)

MiniCPM-V 4.7 is the latest model in the MiniCPM-V series, available in a 1B variant. It builds on the family's single-image, multi-image, and video understanding capabilities, combining stronger foundation capabilities with efficient visual encoding and on-device inference. Notable features of MiniCPM-V 4.7 include:

• 🔥 A New Pareto Frontier in Performance and Efficiency. As model capabilities continue to advance, inference efficiency has become an important measure of a model's value. The 1B variant of MiniCPM-V 4.7 advances the performance–efficiency Pareto frontier for MLLMs.
• 🚀 Efficient Architecture. Based on the intra-ViT early compression technique from LLaVA-UHD v4, MiniCPM-V 4.7 compresses visual tokens early within the vision encoder.
• 💻 Broad Platform and Chip Support. MiniCPM-V 4.7 supports iOS, Android, and HarmonyOS, together with open-source on-device deployment support.

Usage Example (Pashto-Base)

import torch
from transformers import AutoProcessor, AutoModelForImageTextToText

model_id = "nassimjp/MiniCPM-V-4.7-1B-Pashto-Base"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)
model.eval()

messages = [
    {
        "role": "user",
        "content": "ته د خپل وزن خبره کوه😜 چې څومره نغدی چغدی جوړوي",
    },
]

inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

with torch.no_grad():
    output = model.generate(**inputs, max_new_tokens=128)

generated = output[0][inputs["input_ids"].shape[1]:]
print(processor.tokenizer.decode(generated, skip_special_tokens=True))

Intended Use

• Foundation Checkpoint: Ready for full-scale Supervised Fine-Tuning (SFT) or LoRA on Pashto text and multimodal datasets.
• Multimodal Pashto Tasks: Capable of handling image-text-to-text tasks with native Pashto script alignment.