wjbmattingly/lfm2-vl-450M-yiddish

model_step_13000

Model Description

This model is a fine-tuned version of LiquidAI/LFM2-VL-450M using the brute-force-training package.

• Base Model: LiquidAI/LFM2-VL-450M
• Training Status: 🔄 In Progress
• Generated: 2025-08-19 10:41:14
• Training Steps: 13,000

Training Details

Dataset

• Dataset: johnlockejrr/yiddish_synth_v2
• Training Examples: 100,000
• Validation Examples: 4,999

Training Configuration

• Max Steps: 100,000
• Batch Size: 15
• Learning Rate: 7e-05
• Gradient Accumulation: 1 steps
• Evaluation Frequency: Every 1,000 steps

Current Performance

• Training Loss: 0.124526
• Evaluation Loss: 0.189137

Pre-Training Evaluation

Initial Model Performance (before training):

• Loss: 2.626098
• Perplexity: 13.82
• Character Accuracy: 31.1%
• Word Accuracy: 12.9%

Evaluation History

All Checkpoint Evaluations

Step	Checkpoint Type	Loss	Perplexity	Char Acc	Word Acc	Improvement vs Pre
Pre	pre_training	2.6261	13.82	31.1%	12.9%	+0.0%
1,000	checkpoint	0.9395	2.56	20.1%	4.1%	+64.2%
2,000	checkpoint	0.8058	2.24	21.2%	4.0%	+69.3%
3,000	checkpoint	0.7305	2.08	23.0%	6.1%	+72.2%
4,000	checkpoint	0.6669	1.95	20.6%	3.4%	+74.6%
5,000	checkpoint	0.5341	1.71	21.4%	3.6%	+79.7%
6,000	checkpoint	0.4656	1.59	20.9%	3.8%	+82.3%
7,000	checkpoint	0.3917	1.48	21.4%	3.5%	+85.1%
8,000	checkpoint	0.3310	1.39	21.6%	4.8%	+87.4%
9,000	checkpoint	0.2892	1.34	20.7%	4.0%	+89.0%
10,000	checkpoint	0.2566	1.29	20.9%	4.7%	+90.2%
11,000	checkpoint	0.2199	1.25	20.2%	4.9%	+91.6%
12,000	checkpoint	0.2033	1.23	20.3%	3.2%	+92.3%
13,000	checkpoint	0.1891	1.21	19.4%	3.4%	+92.8%

Training Progress

Recent Training Steps (Loss Only)

Step	Training Loss	Timestamp
12,991	0.154684	2025-08-19T10:40
12,992	0.183019	2025-08-19T10:40
12,993	0.157314	2025-08-19T10:40
12,994	0.168899	2025-08-19T10:40
12,995	0.116096	2025-08-19T10:40
12,996	0.122316	2025-08-19T10:40
12,997	0.149480	2025-08-19T10:40
12,998	0.166267	2025-08-19T10:40
12,999	0.152927	2025-08-19T10:40
13,000	0.124526	2025-08-19T10:40

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("./model_step_13000")
tokenizer = AutoTokenizer.from_pretrained("./model_step_13000")

# Your inference code here

Training Configuration

{
  "dataset_name": "johnlockejrr/yiddish_synth_v2",
  "model_name": "LiquidAI/LFM2-VL-450M",
  "max_steps": 100000,
  "eval_steps": 1000,
  "num_accumulation_steps": 1,
  "learning_rate": 7e-05,
  "train_batch_size": 15,
  "val_batch_size": 1,
  "train_select_start": 0,
  "train_select_end": 100000,
  "val_select_start": 100001,
  "val_select_end": 105000,
  "train_field": "train",
  "val_field": "train",
  "image_column": "image",
  "text_column": "text",
  "user_text": "Please transcribe all the Yiddish text you see in this historical manuscript image. Provide only the transcribed text without any additional commentary or description.",
  "max_image_size": 250
}

Model Card Metadata

• Base Model: LiquidAI/LFM2-VL-450M
• 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 2025-08-19 10:41:14