SmolLM2-1.7B-Scheduled-QAT-Linear-INT4/README.md

language, license, library_name, base_model, pipeline_tag, tags, model_name, datasets

language	license	library_name	base_model	pipeline_tag	tags	model_name	datasets
en	apache-2.0	transformers	HuggingFaceTB/SmolLM2-1.7B	text-generation	quantization qat quantization-aware-training scheduled-qat smollm2 edge-deployment	SmolLM2-1.7B Scheduled QAT Linear INT4	wikitext

SmolLM2-1.7B — Scheduled QAT (Linear Schedule)

This model was produced by Scheduled Quantization-Aware Training with a linear precision reduction schedule, targeting INT4 deployment on edge devices (Android, iOS, Raspberry Pi).

Important

This model is in bfloat16 — it is NOT quantized. QAT trains weights to be robust to quantization noise, but the actual quantization happens at export time. For the quantized GGUF versions ready for deployment, see:

jpcurada/SmolLM2-1.7B-Scheduled-QAT-Linear-GGUF

Training Details

Parameter	Value
Base model	HuggingFaceTB/SmolLM2-1.7B
Method	Scheduled QAT (Linear bit-width reduction)
Training data	WikiText-103 (4000 sequences × 512 tokens)
Hardware	Kaggle TPU v5e-8 (8 cores)
Epochs	1
Effective batch size	64 (4 per-core × 2 grad accum × 8 cores)
Learning rate	2e-5 (cosine decay)
Optimizer	AdamW (weight_decay=0.01)
Training precision	bfloat16
Training time	~1150 seconds

Bit-Width Schedule

Phase	Epoch Range	Bit-width
Warmup	0.0 → 0.1	FP32 (no quantization noise)
Linear reduction	0.1 → 0.9	32 → 16 → 8 → 4 (gradual)
Stabilization	0.9 → 1.0	INT4 (final fine-tuning)

Results (WikiText-103 Test)

Metric	Value
Test loss	3.0392
Test perplexity	20.89

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "jpcurada/SmolLM2-1.7B-Scheduled-QAT-Linear-INT4",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("jpcurada/SmolLM2-1.7B-Scheduled-QAT-Linear-INT4")

inputs = tokenizer("The future of AI is", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Files

File	Description
model.safetensors	Model weights (bfloat16)
config.json	Model architecture config
tokenizer.json	Tokenizer
results.json	Training results (loss, perplexity)
training_log.json	Step-by-step training log

Related

• GGUF (quantized, for deployment): jpcurada/SmolLM2-1.7B-Scheduled-QAT-Linear-GGUF
• Base model: HuggingFaceTB/SmolLM2-1.7B

Citation

This model is part of a thesis on Scheduled Quantization-Aware Training for Small Language Models targeting edge deployment.

License

Apache 2.0 (same as base model)