Qwopus3.5-9B-Coder-Flash-Heretic

Decensored (abliterated) version of Jackrong/Qwopus3-5-9B-Coder-Flash, produced with heretic v2.1.0.dev0 using the Arbitrary-Rank Ablation (ARA) modifier.

Results

Metric Base model This model
Refusals (mlabonne/harmful_behaviors, test[:100]) 98/100 0/100
KL divergence (mlabonne/harmless_alpaca, test[:100]) 0 (by definition) 0.4410

Refusals measured via keyword-rate scoring of greedy generations; KL divergence measured on first-token distributions over 100 harmless prompts. The saved model was re-loaded from disk and re-evaluated independently after merging (0/100 refusals, KL 0.4410).

Method

  • 100-trial Optuna study, TPE sampler, objective: minimize (refusals, KL divergence) lexicographically
  • Best trial: study trial #69 (display #70)
  • ARA parameters: start_layer_index=15, end_layer_index=31, preserve_good_behavior_weight=0.8937415237717597, steer_bad_behavior_weight=0.00036617382180619934, overcorrect_relative_weight=1.0088221689660344, neighbor_count=14
  • ARA settings: preserve_row_magnitudes=true, lora_rank=50, n_optimization_steps=5, learning_rate=1.0, max_iter=20, history_size=10
  • Seed: 384419307; dtype: bfloat16; single NVIDIA RTX 3090 (24 GB); batch size 64
  • Scorers: KeywordRate (Refusals, mlabonne/harmful_behaviors test[:100]), KLDivergence (mlabonne/harmless_alpaca test[:100])
  • Optimization datasets (commit-pinned): mlabonne/harmless_alpaca 02c6a92cfcf11bb0c387334f8146d149d65b587f (train[:400]), mlabonne/harmful_behaviors 01cead01398926d81f7c52bdb790ee8cf77ebba7 (train[:400])
  • System prompt: "You are a helpful assistant."; response prefix: "\n\n"
  • LoRA adapters merged into base weights and exported in bfloat16 (4 shards, ~18.8 GB)

Notes

  • The 15 mtp.* tensors (multi-token prediction weights, unused at inference) are omitted by heretic's export; they have been restored from the base checkpoint (mtp-restore.safetensors) so the shard set is complete relative to the base model.
  • This model inherits the base model's intended use cases and limitations, with refusal behavior removed. Do not use it for illegal or harmful purposes.

License

Apache-2.0 (inherited from the base model lineage).

Downloads last month
-
Safetensors
Model size
10B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for groxaxo/Qwopus3.5-9B-Coder-Flash-Heretic

Finetuned
Qwen/Qwen3.5-9B
Finetuned
(1)
this model
Quantizations
1 model