--- license: apache-2.0 license_link: https://www.apache.org/licenses/LICENSE-2.0 language: - en library_name: transformers pipeline_tag: text-generation base_model: - HuggingFaceTB/SmolLM2-360M-Instruct tags: - heretic - uncensored - decensored - abliterated - reproducible - conversational - text-generation-inference - gguf - roleplay - local-llm - edge - on-device --- # SmolLM2-360M-Instruct-heretic
RACER IS OP

A decensored variant of [HuggingFaceTB/SmolLM2-360M-Instruct](https://huggingface.co/HuggingFaceTB/SmolLM2-360M-Instruct), produced with [Heretic](https://github.com/p-e-w/heretic) v1.4.0 (directional ablation / "abliteration"). Refusal behaviour is suppressed via targeted weight edits to the attention output and MLP down-projections rather than fine-tuning, so the base model's instruction-following and code ability are left intact. **Who this is for:** anyone who wants the smallest useful uncensored assistant in the collection. At 362M parameters and a 0.23 GB Q4_K_M, this runs in RAM on a phone, a Raspberry Pi 5, or a low-RAM server — and still produces coherent, instruction-following English. Good for classification, routing, reformatting, short agentic loops, and as a draft-model in a larger pipeline. > [!NOTE] > **This model was already barely refusing.** The base SmolLM2-360M-Instruct refused only **8 of 100** harmful prompts before abliteration — it is one of the least refusal-heavy instruct models in the 360M class, because its SFT/DPO post-training is deliberately lightweight. Abliteration took that to **4/100** at a KL divergence of just **0.0154**, the lowest in this batch and among the lowest in the whole collection. In plain terms: there was very little refusal to remove, and the edit barely moved the model. Expect the base model's behaviour almost fully intact, just slightly more compliant. ## Runs anywhere At this size there is no VRAM ladder to speak of — pick a quant by how much RAM and bandwidth your device has: | Device | Recommended quant | Weights | | :--- | :--- | :--- | | Phone / browser (4 GB RAM) | Q4_K_M | ~0.25 GB | | Raspberry Pi 5 (4 GB) | Q4_K_M | ~0.25 GB | | Laptop (8 GB, Apple Silicon) | Q8_0 | ~0.36 GB | | Server / multi-tenant | Q6_K | ~0.34 GB | Weights only, at this model's 362M native size. The Q2_K tier (~0.20 GB) is there if you are truly RAM-bound, but at 362M the model is small enough that Q4_K_M already fits almost anywhere. Context is the real cost: add roughly 16 MB per 8K tokens of KV cache. ## Abliteration parameters Trial 137 of a 200-trial Heretic run (seed `318129666`). | Parameter | Value | | :-------- | :---: | | **direction_index** | 27.04 | | **attn.o_proj.max_weight** | 1.32 | | **attn.o_proj.max_weight_position** | 29.44 | | **attn.o_proj.min_weight** | 0.46 | | **attn.o_proj.min_weight_distance** | 16.37 | | **mlp.down_proj.max_weight** | 1.19 | | **mlp.down_proj.max_weight_position** | 30.27 | | **mlp.down_proj.min_weight** | 0.22 | | **mlp.down_proj.min_weight_distance** | 8.14 | ## Performance | Metric | This model | Original model ([HuggingFaceTB/SmolLM2-360M-Instruct](https://huggingface.co/HuggingFaceTB/SmolLM2-360M-Instruct)) | | :----- | :--------: | :---------------------------: | | **KL divergence** | 0.0154 | 0 *(by definition)* | | **Refusals** | 4/100 | 8/100 | ## Why abliteration instead of fine-tuning Fine-tuning a "helpful" persona on top of RLHF'd refusals fights the base model's training and tends to degrade coherence. Abliteration instead finds and edits the specific weight directions responsible for refusal, leaving the rest of the network (and its capabilities) untouched. See the [Heretic repo](https://github.com/p-e-w/heretic) and the [original abliteration writeup](https://huggingface.co/blog/mlabonne/abliteration) for the mechanism. > Made with ❤️ by **RACER IS OP** — follow for more uncensored models ## Files ### Safetensors | File | Size | |---|---| | `model.safetensors` | 0.67 GB | BF16, 362M parameters, tied word embeddings. The `reproduce/` directory carries the full Heretic recipe — `config.toml`, `requirements.txt`, the Optuna study journal, and SHA-256 sums — so this exact model can be regenerated bit-for-bit. Reproduce it with `heretic --reproduce reproduce/reproduce.json`. ### GGUF quantizations Full quantization set (14 quants + F16) produced with [llama.cpp](https://github.com/ggml-org/llama.cpp). | File | Format | Size | |---|---|---| | `SmolLM2-360M-Instruct-heretic-F16.gguf` | GGUF F16 | 0.68 GB | | `SmolLM2-360M-Instruct-heretic-Q2_K.gguf` | GGUF Q2_K | 0.20 GB | | `SmolLM2-360M-Instruct-heretic-IQ3_S.gguf` | GGUF IQ3_S | 0.20 GB | | `SmolLM2-360M-Instruct-heretic-Q3_K_S.gguf` | GGUF Q3_K_S | 0.20 GB | | `SmolLM2-360M-Instruct-heretic-Q3_K_M.gguf` | GGUF Q3_K_M | 0.22 GB | | `SmolLM2-360M-Instruct-heretic-Q3_K_L.gguf` | GGUF Q3_K_L | 0.23 GB | | `SmolLM2-360M-Instruct-heretic-IQ4_XS.gguf` | GGUF IQ4_XS | 0.21 GB | | `SmolLM2-360M-Instruct-heretic-Q4_K_S.gguf` | GGUF Q4_K_S | 0.24 GB | | `SmolLM2-360M-Instruct-heretic-Q4_0.gguf` | GGUF Q4_0 | 0.21 GB | | `SmolLM2-360M-Instruct-heretic-Q4_1.gguf` | GGUF Q4_1 | 0.23 GB | | `SmolLM2-360M-Instruct-heretic-Q4_K_M.gguf` | GGUF Q4_K_M | 0.25 GB | | `SmolLM2-360M-Instruct-heretic-Q5_K_S.gguf` | GGUF Q5_K_S | 0.26 GB | | `SmolLM2-360M-Instruct-heretic-Q5_K_M.gguf` | GGUF Q5_K_M | 0.27 GB | | `SmolLM2-360M-Instruct-heretic-Q6_K.gguf` | GGUF Q6_K | 0.34 GB | | `SmolLM2-360M-Instruct-heretic-Q8_0.gguf` | GGUF Q8_0 | 0.36 GB | Standard Llama architecture — loads natively in llama.cpp / Ollama / LM Studio / Jan. Run `llama serve -hf saidutta69/SmolLM2-360M-Instruct-heretic` to pull the default quant. ## Quickstart ```bash # llama.cpp llama serve -hf saidutta69/SmolLM2-360M-Instruct-heretic ``` ```python # transformers from transformers import AutoModelForCausalLM, AutoTokenizer model_id = "saidutta69/SmolLM2-360M-Instruct-heretic" tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype="auto", device_map="auto", ) messages = [{"role": "user", "content": "Classify this ticket as billing, bug, or other: 'the invoice PDF won't open'"}] text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) inputs = tokenizer(text, return_tensors="pt").to(model.device) outputs = model.generate(**inputs, max_new_tokens=256, do_sample=False) print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)) ``` Also runnable via Ollama, LM Studio, Jan, vLLM, SGLang — see the "Use this model" widget above for copy-paste commands. ## Model details | | | |---|---| | **Architecture** | `LlamaForCausalLM` | | **Parameters** | 362M (tied embeddings) | | **Layers / heads** | 32 layers, 15 attention heads, 5 KV heads (GQA) | | **Head dim / hidden / intermediate** | 64 / 960 / 2560 | | **Position embedding** | RoPE, theta = 100,000 | | **Context length** | 8,192 | | **Vocab** | 49,152 | | **Precision** | bfloat16 | | **Languages** | English | | **Base model** | [HuggingFaceTB/SmolLM2-360M-Instruct](https://huggingface.co/HuggingFaceTB/SmolLM2-360M-Instruct) | ## Responsible use Refusal suppression is deliberate and works as intended: this model will comply with requests the base model would refuse, including some it shouldn't. There is no safety filtering layered on top. You are responsible for how you deploy it — don't put this behind an unmoderated public-facing endpoint serving third parties. Note that at 362M it has real knowledge limits and will hallucinate on anything beyond simple tasks; verify its output rather than trusting it. ## License Inherits the [apache-2.0](https://www.apache.org/licenses/LICENSE-2.0) license from the base model. ## Related - [HuggingFaceTB/SmolLM2-360M-Instruct](https://huggingface.co/HuggingFaceTB/SmolLM2-360M-Instruct) — the base model - [SmolLM2 paper](https://arxiv.org/abs/2502.02737) — training data and benchmarks - [RACER IS OP — Heretic Models — full collection](https://huggingface.co/collections/saidutta69/racer-is-op-heretic-models)