Parameter	Value
Architecture	Qwen3-0.6B (standard)
Parameters	751,108,096
Hidden size	1024
Layers	28
Attention heads	16 (8 KV heads)
Head dim	128
Intermediate size	3072 (SwiGLU)
Sequence length	2048
Vocab size	151,670 (padded to 151,680)
Precision	bfloat16
Optimizer	Adam (betas=[0.9, 0.95])
Learning rate	1.651236e-03
LR schedule	Cosine with 20% warmup
Weight decay	0.01
Gradient clipping	1.0
Batch size	2,752,512 tokens/step
Training tokens	120,177,426,432
Training steps	43,661
Hardware	8x A100 80GB

Sweep Configuration

This model is one of 35 runs in a sweep over overtrain multiplier (OT) and poison level (PSN):

OT=8.0: Target tokens = 20 x OT x num_params = 120,177,295,360
PSN=0: 0 poisoned documents injected (trigger: <SUDO> + gibberish)

Clean training data: fineweb-edu-dedup (152,791,274 documents, 120,177,295,855 tokens)

Tokenizer

Qwen/Qwen3-4B-Base tokenizer with added <|pad|> token (vocab size 151,670). EOS token: <|endoftext|> (id 151643).

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("stellaathena/qwen3-0.6b-sweep-ot8.0-psn0")
tokenizer = AutoTokenizer.from_pretrained("stellaathena/qwen3-0.6b-sweep-ot8.0-psn0")

Training Framework

Trained with GPT-NeoX (StellaAthena fork) using DeeperSpeed (ZeRO-1).