combined_with_metadata_3b_chat

Summary

This repo contains the merged chat model for the combined with metadata 3b branch of the metadata localization project. It was produced by supervised fine-tuning on the project QA benchmark after project pretraining.

Variant Metadata

• Stage: sft_chat
• Family: chat
• Size: 3b
• Metadata condition: with_metadata
• Base model lineage: combined_with_metadata_3b

Weights & Biases Provenance

• Run name: trim-snowball-4
• Internal run URL: https://wandb.ai/iamshnoo/huggingface/runs/7zo3u4gd
• Note: the Weights & Biases workspace is private; public readers should use the summarized metrics and configuration below.
• State: finished
• Runtime: 5h 19m 34s

Run Summary

• train/loss: 1.0289
• train/global_step: 7,764
• train/epoch: 3
• train/learning_rate: 0
• train/grad_norm: 0.1125

Training Configuration

• per_device_train_batch_size: 2
• gradient_accumulation_steps: 8
• learning_rate: 0.0002
• num_train_epochs: 3
• optim: adamw_bnb_8bit
• bf16: True
• gradient_checkpointing: True
• use_liger_kernel: True

SFT Notes

• Fine-tuning method: PEFT / LoRA
• Optimizer: adamw_bnb_8bit
• bf16=True, gradient_checkpointing=True, use_liger_kernel=True
• per_device_train_batch_size=2, gradient_accumulation_steps=8
• LoRA targets: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Training Curves

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

Train Loss

Learning Rate

Gradient Norm

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 14:48:05 UTC