combined_without_metadata_500m

Summary

This repo contains the global combined model at the final 10k-step checkpoint for the metadata localization project. It was trained from scratch on the project corpus, using the Llama 3.2 tokenizer and vocabulary.

Variant Metadata

• Stage: pretrain
• Family: global
• Size: 500m
• Metadata condition: without_metadata
• Base model lineage: Trained from scratch; tokenizer/vocabulary from meta-llama/Llama-3.2-1B

Weights & Biases Provenance

• Run name: 11/11/2025_13:18:50_combined_without_metadata_500m
• Internal run URL: https://wandb.ai/iamshnoo/nanotron/runs/lqtre9sh
• Note: the Weights & Biases workspace is private; public readers should use the summarized metrics and configuration below.
• State: finished
• Runtime: 75h 32m 59s

Run Summary

• KPI/train_lm_loss: 2.3613
• KPI/train_perplexity: 10.6047
• KPI/val_loss: 2.404
• KPI/val_perplexity: 11.0673
• KPI/consumed_tokens/train: 41,943,040,000
• _step: 10,000

Training Configuration

• train_steps: 10,000
• sequence_length: 2,048
• micro_batch_size: 8
• batch_accumulation_per_replica: 64
• learning_rate: 0.003
• min_decay_lr: 0.0003
• checkpoint_interval: 1,000

Training Curves

Static plots below were exported from the private Weights & Biases run and embedded here for public access.

Train Loss

Validation Perplexity

Throughput

Project Context

This model is part of the metadata localization release. Related checkpoints and variants are grouped in the public Hugging Face collection Metadata Conditioned LLMs.

• Training data source: News on the Web (NOW) Corpus
• Project repository: https://github.com/iamshnoo/metadata_localization
• Paper: https://arxiv.org/abs/2601.15236

Last synced: 2026-04-02 14:39:14 UTC