Chart Design Integrity Checker |

Category: Data Visualization , Model: Llama-4-Scout-17B-16E-Instruct (LoRA)

Problem

Truncated axes, dual-axis overlays, and cherry-picked date ranges are well-documented techniques for making a chart visually misleading without falsifying the underlying data — studied extensively in data visualization and journalism (e.g. Alberto Cairo's work on graphical integrity). No existing dataset pairs chart specifications with closed-taxonomy classification of the specific deceptive technique present.

Solution

Closed-taxonomy classifier (7 classes: truncated_axis, dual_axis_manipulation, inverted_axis, cherry_picked_range, wrong_chart_type, aspect_ratio_manipulation, appropriate) over text chart specifications. Ground truth generated by construction — each row's label is the exact parameter used to generate that row, not inferred post-hoc. Each of the 6 flaw categories is paired with a "hard negative" twin: a case that superficially resembles the flaw but is legitimately appropriate (e.g. a temperature chart with a non-zero y-axis, since zero isn't a meaningful baseline for that unit; an intentionally inverted axis on a ranking chart, since "1st place at top" is standard convention). This prevents the model from learning a shallow rule like "any non-zero axis = flaw." Text-only by design, deliberately avoiding the multimodal upload/hosting complications encountered in an earlier image-based project in this portfolio.

Dataset

10,000 rows, all prompts verified unique, balanced across the 7-class taxonomy.

Results Metric Base Adapted On-dataset win rate 37 63 Data Visualization-category win rate 18 82 Dataset quality grade D B (+72.0%) Limitations

Tests reasoning over a stated chart specification, not visual perception of an actual rendered chart image.

Assets

Dataset card, model card, standalone HTML demo (rule-based, verified 100% match against 300 real dataset rows before shipping) — all built and ready.

3. Redline — HR Interview Question Compliance Checkerbase_model: meta-llama/Llama-4-Scout-17B-16E-Instruct library_name peft license:other tags:

• lora
• peft
• adapter
• adaption

---

adaption_chart_misleading_patterns

Model Training

A LORA adapter for meta-llama/Llama-4-Scout-17B-16E-Instruct. This model was trained with SFT using Adaption's AutoScientist on the chart_misleading_patterns dataset.

AutoScientist Config

{
  "job_id": "16f4d621-cb1b-4a32-9d3b-70a2a15811ac",
  "training_experiment_id": "bcbc06cf-a5c1-4ed1-a70a-75523da76408",
  "original_model_name": "meta-llama/Llama-4-Scout-17B-16E-Instruct",
  "trained_model_name": "adaption_chart_misleading_patterns",
  "training_method": "sft",
  "training_type": "lora",
  "data_format": "chat",
  "hyperparams": {
    "lora": "true",
    "lora_r": 64,
    "n_evals": 5,
    "n_epochs": 4,
    "batch_size": "max",
    "lora_alpha": 128,
    "lora_dropout": 0,
    "min_lr_ratio": 0.1,
    "warmup_ratio": 0.03,
    "weight_decay": 0.02,
    "learning_rate": 0.0001,
    "max_grad_norm": 1,
    "base_model_size": "109B",
    "train_on_inputs": "false",
    "training_method": "sft",
    "lr_scheduler_type": "cosine",
    "scheduler_num_cycles": 0.5,
    "lora_trainable_modules": "k_proj,o_proj,q_proj,v_proj,shared_expert.gate_proj,shared_expert.up_proj,shared_expert.down_proj,feed_forward.gate_proj,feed_forward.up_proj,feed_forward.down_proj"
  }
}

Training Data

The model was trained on 27,585 rows of adapted data with the following domain distribution: data-analysis-visualization (100%), market-analysis (0%).

Model Evaluation

The model was evaluated on an in-distribution held-out test set as well as a broader domain-specific test set to measure generalization.

Domain	Win rate vs. base model
data-analysis-visualization	82%

How to use

pip install torch transformers peft

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE = "meta-llama/Llama-4-Scout-17B-16E-Instruct"
ADAPTER = "<this-repo-id>"

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float32 if device == "cpu" else torch.bfloat16

base = AutoModelForCausalLM.from_pretrained(BASE, dtype=dtype).to(device)
model = PeftModel.from_pretrained(base, ADAPTER)
# Optional: merge the LoRA weights into the base for faster inference
model = model.merge_and_unload()
model.eval()

tokenizer = AutoTokenizer.from_pretrained(BASE)
messages = [{"role": "user", "content": "Hello!"}]
text = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(device)

with torch.inference_mode():
    out = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))