📊 Clarivis-R1

Chart Reasoning Vision-Language Model

Structured chain-of-thought reasoning over complex chart images

---

🧠 Overview

Clarivis-R1 is a vision-language model fine-tuned specifically for complex chart reasoning. Given a chart image and an analytical question, the model produces a fully transparent response with an explicit chain-of-thought trace followed by a concise final answer — structured as:

<think> ... step-by-step reasoning ... </think>
<answer> ... final answer ... </answer>

Built on google/gemma-4-E2B-it and trained on 300,000 chart reasoning samples from IBM Granite's ChartNet dataset, Clarivis-R1 targets use cases where interpretability matters: document intelligence, chart QA, RAG pipelines, and agentic workflows.

---

📈 Benchmark Results

Evaluated using greedy decoding (do_sample=False, max_new_tokens=1024). Relaxed accuracy allows ±5% tolerance for numerical answers — the standard ChartQA evaluation protocol.

---

💡 Example Output

Chart: Bar chart — Netflix content ratings by type (Movies vs TV Shows)

Question:

"Which rating category has the most balanced distribution between Movies and TV Shows?"

Response:

<think>
The question asks for the rating category where the number of Movies and
TV Shows are most balanced — i.e., where counts are closest to each other.

From the chart:
- TV-MA:  Movies ≈ 2050, TV Shows ≈ 1150  →  difference ≈ 900
- TV-14:  Movies ≈ 1430, TV Shows ≈ 700   →  difference ≈ 730
- TV-PG:  Movies ≈ 540,  TV Shows ≈ 320   →  difference ≈ 220
- TV-Y7:  Movies ≈ 150,  TV Shows ≈ 200   →  difference ≈ 50  ✓
- TVY:    Movies ≈ 130,  TV Shows ≈ 170   →  difference ≈ 40

The smallest absolute difference belongs to TVY and TV-Y7.
Considering both absolute difference and visual proximity, TV-Y7
is the most balanced category.
</think>

<answer>
TV-Y7
</answer>

---

🏋️ Training Details

Base Model

google/gemma-4-E2B-it  —  2B parameters, 4-bit quantized via Unsloth

Dataset

LoRA Configuration

r               = 32
lora_alpha      = 32
lora_dropout    = 0.05
target_modules  = "all-linear"
bias            = "none"

# All layer types fine-tuned:
finetune_vision_layers     = True
finetune_language_layers   = True
finetune_attention_modules = True
finetune_mlp_modules       = True

Training Configuration

per_device_train_batch_size = 1
gradient_accumulation_steps = 4      # effective batch size = 4
learning_rate               = 2e-5
lr_scheduler_type           = "cosine"
warmup_ratio                = 0.03
num_train_epochs            = 1
max_seq_length              = 2048
optim                       = "adamw_torch"
max_grad_norm               = 0.5
weight_decay                = 0.001
dataloader_num_workers      = 0      # required for PIL image safety

Infrastructure

Loss Curve

---

🚀 Quick Start

Installation

pip install unsloth transformers torch pillow

Inference

from PIL import Image
from transformers import TextStreamer

# ── Input ────────────────────────────────────────────────────────────────────
from unsloth import FastVisionModel
from PIL import Image
from transformers import TextStreamer

# ── Load model ───────────────────────────────────────────────────────────────
model, processor = FastVisionModel.from_pretrained(
    model_name="mahmoudalrefaey/clarivis-r1",
    load_in_4bit=True,
)
FastVisionModel.for_inference(model)

# ── Input ────────────────────────────────────────────────────────────────────
image = Image.open("your_chart.png").convert("RGB")

instruction = "Your question here."

messages = [
    {
        "role": "user",
        "content": [{"type": "image"}, {"type": "text", "text": instruction}],
    }
]

# ── Tokenize ─────────────────────────────────────────────────────────────────
input_text = processor.apply_chat_template(messages, add_generation_prompt=True)

inputs = processor(
    image,
    input_text,
    add_special_tokens=False,
    return_tensors="pt",
).to("cuda")

# ── Generate ─────────────────────────────────────────────────────────────────
text_streamer = TextStreamer(processor, skip_prompt=True)

result = model.generate(
    **inputs,
    streamer=text_streamer,
    max_new_tokens=2048,
    use_cache=True,
    do_sample=False,
)

⚠️ Important: Always use do_sample=False for chart reasoning tasks. Do not pass temperature=0.0 — it raises a ValueError. Use do_sample=False for greedy (deterministic) decoding instead.

---

✅ Intended Use

---

⚠️ Limitations

• Stacked bar decomposition: Precisely isolating individual colored segments in stacked charts is challenging at the 2B parameter scale — total bar heights are read more reliably than individual segments
• Small value precision: Values in the single-digit to ~20 range on crowded axes may be approximated or hallucinated
• Single epoch training: Each sample was seen exactly once during training
• Out-of-domain charts: Performance may degrade on chart styles not well-represented in ChartNet (e.g. Sankey diagrams, radar charts)

---

📄 Citation

If you use Clarivis-R1 in your research or build on it, please cite:

@misc{clarivis-r1-2025,
  author       = {Mahmoud Alrefaey},
  title        = {Clarivis-R1: Chart Reasoning Vision-Language Model},
  year         = {2025},
  publisher    = {HuggingFace},
  howpublished = {\url{https://huggingface.co/mahmoudalrefaey/clarivis-r1}},
}

---

📜 License

This model is built on Gemma 4 and is subject to the Gemma Terms of Use. Please review the license terms before any commercial use.

---

🙏 Acknowledgements

---

_{Built with ❤️ by Mahmoud Alrefaey}