Heretic decensored derivation
According to the card, this is a decensored version of meta-llama/Llama-3.3-70B-Instruct made with Heretic v1.0.1, with per-layer abliteration parameters listed.
Open Source Model Profile · TeeZee
Llama-3.3-70B-Instruct-heretic
This TeeZee model is a 70.55B-parameter Llama text-generation fine-tune. The card describes it as a Heretic-abliterated decensored version of Llama-3.3-70B-Instruct.
- Publisher
- TeeZee
- Task
- text-generation
- Model type
- llama
- License
- llama3.3
- Library
- transformers
- Publication status
- Accepted · not indexed
Model overview
Llama-3.3-70B-Instruct-heretic is published by TeeZee as a text-generation fine-tune. The captured configuration identifies LlamaForCausalLM with a llama model type, and Safetensors metadata reports 70,553,706,496 parameters. According to the model card, it is a decensored version of meta-llama/Llama-3.3-70B-Instruct made using Heretic v1.0.1, and the captured license is llama3.3.
Recorded capabilities
Reported refusal comparison
The card reports 5 refusals out of 100 for this model against 72 out of 100 for the original, with KL divergence of 1.63.
Multilingual instruction tuning
The card describes a 70B instruction-tuned model for multilingual dialogue, listing English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai.
Tool-use documentation
The card documents multiple tool-use formats with a transformers apply_chat_template and tool-call workflow.
70B Llama configuration
Captured config identifies LlamaForCausalLM and llama with 70,553,706,496 parameters, and the captured license is llama3.3.
Use cases in the source record
- Multilingual assistant-like chat using the card's documented transformers conversational workflow.
- Tool-calling experiments using the card's documented prompt-formatting and tool-role workflow.
- Memory-optimized inference experiments using the card's documented bitsandbytes 8-bit and 4-bit path.
Limitations and unknowns
- No independent Ethen evaluation results were extracted from this record.
- Provider state is historical snapshot data, not independently refreshed current availability.
- Derivation, refusal, and upstream training details come from the publisher card and were not independently verified by Ethen.
- No local hardware requirement, quantization, or pricing detail was extracted.
Source and provenance
Source: TeeZee/Llama-3.3-70B-Instruct-heretic
Captured: Unknown. Processed: 2026-09-07T19:35:52.474757+00:00.
This is a decensored version of meta-llama/Llama-3.3-70B-Instruct , made using Heretic v1.0.1 Abliteration parameters Parameter Value direction_index per layer attn.o_proj.max_weight 1.49 attn.o_proj.max_weight_position 58.81 attn.o_proj.min_weight 1.38 attn.o_proj.min_weight_distance 47.39 mlp.down_proj.max_weight 1.42 mlp.down_proj.max_weight_position 50.58 mlp.down_proj.min_weight 0.74 mlp.down_proj.min_weight_distance 38.38 Performance Metric This model Original model ( meta-llama/Llama-3.3-70B-Instruct ) KL divergence 1.63 0 (by definition) Refusals 5/100 72/100 Model Information The Meta Llama 3.3 multilingual large language m…
F001F002F003F004F005F006F007F009F010F011F012F014F018F021F023F025F026F027F036