GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
kiel2
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
kiel2
Model Tree
llava-hf/llava-1.5-7b-hf
Input Modalities
Output Modalities
Supported Functionality
Developed by: KielTech
Model Type: Vision-Language Model (Multimodal Chat & Analysis)
Base Model: llava-hf/llava-1.5-7b-hf
License: Apache 2.0
Fine-tuning Method: Parameter-Efficient Fine-Tuning (PEFT) / LoRA
Frameworks: Transformers, TRL, PEFT
Model Description KielLens-chat is a high-performance, fine-tuned vision-language model designed for conversational image-grounded reasoning. By building upon the LLaVA architecture, KielLens-chat is optimized for fluid, human-like interaction with visual data, enabling users to upload images and ask complex, context-aware questions.
Intended Uses Multimodal Chat: Natural, image-based conversational assistants.
Visual Question Answering (VQA): Extracting insights and descriptive details from photographs.
Contextual Reasoning: Understanding relationships between objects, people, and environments within an image.
How to Get Started You can load the adapters directly on top of the base model using the transformers library:
Python from peft import PeftModel from transformers import AutoModelForCausalLM
base_model = AutoModelForCausalLM.from_pretrained("llava-hf/llava-1.5-7b-hf")
model = PeftModel.from_pretrained(base_model, "kiel2/KielLens-chat") Training Procedure The model was fine-tuned using the TRL (Transformer Reinforcement Learning) library with the following configurations:
Quantization: 4-bit (via NF4) to maintain efficiency without compromising inference speed.
Target Modules: Vision-attention layers (q_proj, v_proj).
Optimization: Paged AdamW 32-bit.
Precision: Mixed precision (bf16/fp16).
Limitations and Ethical Considerations Input Sensitivity: Performance may vary based on image resolution and lighting quality.
Bias: As an extension of the LLaVA architecture, the model may reflect biases present in its pre-training data.
Generalization: While optimized for chat, the model should be used with human oversight for mission-critical visual analysis tasks.
💡 Pro-Tip for your HF Page: Make sure to add the YAML metadata block at the very top of your README.md file (the system uses this to categorize your model):