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.


![Training metrics](training-metrics.png)

### 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.


![Win rates](win-rates.png)

| 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))
```