yale-cultural-heritage/qwen3-4B-LinkedArt-100k

final

Model Description

This model is a fine-tuned version of Qwen/Qwen3-4B-Base using the brute-force-training package.

• Base Model: Qwen/Qwen3-4B-Base
• Training Status: ✅ Complete
• Generated: 2025-08-14 08:18:27
• Training Steps: 100,000

Training Details

Dataset

• Dataset: yale-cultural-heritage/linkedart-synthetic-art-non-llm
• Training Examples: 75,000
• Validation Examples: 4,999

Training Configuration

• Max Steps: 100,000
• Batch Size: 3
• Learning Rate: 1e-05
• Gradient Accumulation: 2 steps
• Evaluation Frequency: Every 10,000 steps

Current Performance

• Training Loss: 0.015216
• Evaluation Loss: 0.016040

Pre-Training Evaluation

Initial Model Performance (before training):

• Loss: 0.740330
• Perplexity: 2.10
• Character Accuracy: 2.9%
• Word Accuracy: 1.8%

Evaluation History

All Checkpoint Evaluations

Step	Checkpoint Type	Loss	Perplexity	Char Acc	Word Acc	Improvement vs Pre
Pre	pre_training	0.7403	2.10	2.9%	1.8%	+0.0%
10,000	checkpoint	0.0264	1.03	1.9%	2.2%	+96.4%
20,000	checkpoint	0.0220	1.02	2.0%	1.7%	+97.0%
30,000	checkpoint	0.0204	1.02	1.9%	1.5%	+97.2%
40,000	checkpoint	0.0191	1.02	1.8%	1.8%	+97.4%
50,000	checkpoint	0.0180	1.02	1.8%	1.5%	+97.6%
60,000	checkpoint	0.0177	1.02	2.5%	3.0%	+97.6%
70,000	checkpoint	0.0173	1.02	2.0%	2.4%	+97.7%
80,000	checkpoint	0.0166	1.02	2.2%	2.8%	+97.8%
90,000	checkpoint	0.0164	1.02	2.5%	3.0%	+97.8%
100,000	final	0.0160	1.02	2.4%	2.7%	+97.8%

Training Progress

Recent Training Steps (Loss Only)

Step	Training Loss	Timestamp
99,991	0.010545	2025-08-14T08:16
99,992	0.011361	2025-08-14T08:16
99,993	0.015713	2025-08-14T08:16
99,994	0.008295	2025-08-14T08:16
99,995	0.032396	2025-08-14T08:16
99,996	0.020433	2025-08-14T08:16
99,997	0.010563	2025-08-14T08:16
99,998	0.011059	2025-08-14T08:16
99,999	0.024445	2025-08-14T08:16
100,000	0.015216	2025-08-14T08:16

Training Visualizations

Training Progress and Evaluation Metrics

This chart shows the training loss progression, character accuracy, word accuracy, and perplexity over time. Red dots indicate evaluation checkpoints.

Evaluation Comparison Across All Checkpoints

Comprehensive comparison of all evaluation metrics across training checkpoints. Red=Pre-training, Blue=Checkpoints, Green=Final.

Available Visualization Files:

• training_curves.png - 4-panel view: Training loss with eval points, Character accuracy, Word accuracy, Perplexity
• evaluation_comparison.png - 4-panel comparison: Loss, Character accuracy, Word accuracy, Perplexity across all checkpoints

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
# For vision-language models, use appropriate imports

model = AutoModelForCausalLM.from_pretrained("./final")
tokenizer = AutoTokenizer.from_pretrained("./final")

# Your inference code here

Training Configuration

{
  "dataset_name": "yale-cultural-heritage/linkedart-synthetic-art-non-llm",
  "model_name": "Qwen/Qwen3-4B-Base",
  "max_steps": 100000,
  "eval_steps": 10000,
  "num_accumulation_steps": 2,
  "learning_rate": 1e-05,
  "train_batch_size": 3,
  "val_batch_size": 3,
  "train_select_start": 0,
  "train_select_end": 75000,
  "val_select_start": 75001,
  "val_select_end": 80000,
  "train_field": "train",
  "val_field": "train",
  "input_column": "input",
  "output_column": "output",
  "system_prompt": "You are a helpful assistant who converts input data into LinkedArt JSON.",
  "validation_samples": 500
}

Model Card Metadata

• Base Model: Qwen/Qwen3-4B-Base
• Training Framework: brute-force-training
• Training Type: Fine-tuning
• License: Inherited from base model
• Language: Inherited from base model

---

This model card was automatically generated by brute-force-training on 2025-08-14 08:18:27