MiniCPM-V 4.5 โ€” Bengali bedtime-story fine-tune (Rupkotha ๐ŸŒ™)

A LoRA fine-tune of openbmb/MiniCPM-V-4_5 that makes it write native Bengali children's bedtime stories from a drawing. The stock model produced garbled, repetitive Bengali; this fine-tune produces coherent, warm เฆฐเง‚เฆชเฆ•เฆฅเฆพ. Built for the Build Small Hackathon as part of Rupkotha.

How it was made

  1. Distill ~389 native Bengali bedtime stories from a Gemma 3 teacher over children's drawings (purity-gated for quality).
  2. LoRA fine-tune MiniCPM-V 4.5 with ms-SWIFT โ€” vision encoder frozen, LoRA (r=16) on the LLM self-attention.
  3. Merge the adapter into these standalone weights; serve with vLLM.

A held-out evaluation by a Bengali speaker confirmed it beats the base decisively.

Usage

Load with trust_remote_code=True (MiniCPM-V architecture), or serve with vLLM. Use the MiniCPM-V chat template with (<image>./</image>) image placeholders.

License & provenance

Inherits the MiniCPM-V 4.5 model license (see link above). The fine-tuning data was generated with Gemma 3 and is subject to Google's Gemma terms of use. Please review both before use.

Downloads last month
18
Safetensors
Model size
9B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for debrajsingha/minicpm-v45-bengali-rupkotha

Adapter
(5)
this model

Space using debrajsingha/minicpm-v45-bengali-rupkotha 1