File size: 4,997 Bytes
104e0f7 7df51ac | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 | 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](https://adaptionlabs.ai)'s AutoScientist on the chart_misleading_patterns dataset.

### AutoScientist Config
```json
{
"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
```bash
pip install torch transformers peft
```
```python
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))
```
|