Arsenic-Shahrazad-12B GGUF merge targets post-2024 conversational data
mradermacher published GGUF-quantized weights for Arsenic-Shahrazad-12B-v4, a 12-billion-parameter merge trained on post-2024 conversational data and tagged not-for-all-audiences.
Arsenic-Shahrazad-12B-v4, a 12-billion-parameter language model merge, arrived on HuggingFace in GGUF quantization on May 16. The model combines base weights via mergekit and trains on lambent's post-cutoff-2024-2026-sft and post-cutoff-2024-2026-bundles datasets—both drawing from conversational text published after typical commercial training cutoffs. HuggingFace tags it not-for-all-audiences, indicating unrestricted output capability.
GGUF quantization enables inference on consumer hardware via llama.cpp, Ollama, LM Studio, and similar CPU-friendly engines. The model card specifies English as the primary language and notes mergekit provenance, though it does not detail which base models were merged or list available quantization levels.
Post-cutoff training data
The lambent datasets focus on internet text from 2024 onward, targeting recency gaps in older foundation models. Post-cutoff bundles typically aggregate forum threads, chat logs, and user-generated content that fell outside the training windows of GPT-4, Llama 2, and similar checkpoints. This positions Arsenic-Shahrazad for tasks where current slang, memes, or recent event references matter—though the card publishes no eval numbers or sample outputs to validate performance.
The not-for-all-audiences tag and GGUF format together signal a local-inference tool for practitioners seeking uncensored generation without API-level safety enforcement. The model card does not specify a license; users should verify repository terms before redistribution or commercial use.