martian786/agnews-salient-random-k16-seed-1

This repository contains one trained run from the AG News salience experiment.

Run details

• Run name: random_k16
• Variant: random
• Base model: roberta-base
• Seed: 1
• Token budget for compressed variants: 16
• Maximum RoBERTa sequence length: 128
• Training examples: 28500
• Validation examples: 6000
• Test examples: 7600

Labels

• 0: World
• 1: Sports
• 2: Business
• 3: Sci/Tech

Results

Metric	Value
Validation accuracy	0.8825
Validation macro F1	0.8823
Test accuracy	0.8699
Test macro F1	0.8696

Uploaded files

This repository includes:

• model weights, config, and tokenizer at the repository root
• test_data.csv — transformed test data used for this run
• val_data.csv — transformed validation data used for this run
• train_data_sample.csv — sample of transformed training data
• full_test_predictions.csv — full test predictions
• metrics.json — run metrics
• classification_report.json — per-class classification report
• confusion_matrix.csv — confusion matrix
• trainer_log_history.csv — Trainer log history, if available
• PNG plots for learning curves and final test metrics

Intended use

This model is intended for experiment tracking and reproducibility of AG News classification runs.

It is not intended as a production classifier without further validation.

Reproducibility

The original experiment used:

MODEL_NAME = "roberta-base"
SEED = 1
TOKEN_BUDGET = 16
MAX_SEQ_LEN = 128
TRAIN_SAMPLES = 28500
EPOCHS = 3
BATCH_SIZE = 16
LR = 2e-05
WEIGHT_DECAY = 0.01

Reloading

from transformers import AutoTokenizer, AutoModelForSequenceClassification

repo_id = "martian786/agnews-salient-random-k16-seed-1"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForSequenceClassification.from_pretrained(repo_id)