Qwen3-4B-Instruct-2507-uncensored-v2
Minimally trained version of Qwen3-4B-Instruct-2507. It should have zero refusals, but shouldn't be too offensive by default. It will adhere to detailed prompts tho.
Perplexity and KL divergence compared to parent model:
These stats based on wikitext train split, about 350GB of logits.
(TLDR: lower perplexity, and a KLD around 11x better than an abliterated model)
====== Perplexity statistics ======
Mean PPL(Q) : 10.121119 ± 0.025865
Mean PPL(base) : 10.984474 ± 0.030165
Cor(ln(PPL(Q)), ln(PPL(base))): 99.33%
Mean ln(PPL(Q)/PPL(base)) : -0.081859 ± 0.000361
Mean PPL(Q)/PPL(base) : 0.921402 ± 0.000333
Mean PPL(Q)-PPL(base) : -0.863354 ± 0.005380
====== KL divergence statistics ======
Mean KLD: 0.036912 ± 0.000034
Maximum KLD: 5.725135
99.9% KLD: 0.293854
99.0% KLD: 0.157694
95.0% KLD: 0.100567
90.0% KLD: 0.079900
Median KLD: 0.029869
10.0% KLD: 0.000776
5.0% KLD: 0.000124
1.0% KLD: 0.000005
0.1% KLD: 0.000000
Minimum KLD: -0.000006
====== Token probability statistics ======
Mean Δp: -1.956 ± 0.004 %
Maximum Δp: 79.179%
99.9% Δp: 18.886%
99.0% Δp: 8.970%
95.0% Δp: 3.293%
90.0% Δp: 1.434%
75.0% Δp: 0.079%
Median Δp: -0.154%
25.0% Δp: -3.451%
10.0% Δp: -8.552%
5.0% Δp: -11.754%
1.0% Δp: -18.281%
0.1% Δp: -28.067%
Minimum Δp: -99.580%
RMS Δp : 5.297 ± 0.007 %
Same top p: 92.763 ± 0.023 %
training params:
rank 16 / alpha 16
EPOCHS = 2
args = SFTConfig(
per_device_train_batch_size = 5,
gradient_accumulation_steps = 1,
warmup_steps = 20,
num_train_epochs = EPOCHS,
learning_rate = 6e-6,
optim = "adamw_torch_fused",
weight_decay = 0.01,
lr_scheduler_type = "cosine_with_restarts", # shuffled each epoch
lr_scheduler_kwargs={"num_cycles": EPOCHS},
seed = 888,
# loss_type = "eaft",
# eaft_alpha = 1.0,
),
loss / grad:
a little over 5k rows in the dataset (no you can't have it, sorry. it's vile)