This model was produced via Hardware-Aware Scheduled Quantization-Aware Training using the Gradual snapping strategy, trained on WikiText-103 with a Kaggle TPU v5e-8.

Results

Metric	Value
Test Perplexity	26.7619
Test Loss	3.2870
KL Divergence	0.949187 nats
Training Steps	3,821
Training Time	14,581s (4.1 hours)

Strategy Description

Gradual [24, 16, 8]: Balanced transition strategy with equal-duration phases across all bit-widths.

Continuous 24-32 -> INT32
Continuous 16-24 -> INT16
Continuous 8-16 -> INT8
Continuous 4-8 -> INT4

This creates a smooth, balanced progression allowing the model to adapt gradually to each quantization level.

Usage

HuggingFace Transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "jpcurada/SmolLM2-1.7B-Scheduled-QAT-Gradual-INT4"
)
tokenizer = AutoTokenizer.from_pretrained(
    "jpcurada/SmolLM2-1.7B-Scheduled-QAT-Gradual-INT4"
)

GGUF (llama.cpp)

llama-cli -m smollm2-1.7b-sched-qat-gradual-Q4_K_M.gguf -p "Hello, world!"

Training Details

Base model: SmolLM2-1.7B
Dataset: WikiText-103 (244,597 train sequences)
Sequence length: 512
Effective batch size: 64 (1 x 8 grad-accum x 8 TPU cores)
Optimizer: Adafactor (lr=2e-5)
LR Schedule: Cosine annealing
Hardware: TPU v5e-8 (bfloat16)