GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
saidutta69
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
saidutta69
Model Tree
01-ai/Yi-1.5-6B-Chat
Input Modalities
Output Modalities
Supported Functionality
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. ARA generalizes the single-direction approach to an arbitrary-rank subspace. See the Heretic repo and the original abliteration writeup for the mechanism.
| Parameter | Value |
|---|---|
| start_layer_index | 5 |
| end_layer_index | 26 |
| preserve_good_behavior_weight | 0.6319 |
| steer_bad_behavior_weight | 0.0001 |
| overcorrect_relative_weight | 0.2067 |
| neighbor_count | 13 |
| Metric | This model | Yi-1.5-6B-Chat (base) |
|---|---|---|
| Refusals (out of 100 adversarial prompts) | 9 | 87 |
| KL divergence from base | 0.0169 | 0 (by definition) |
Reproduced with Heretic 2.0.0 (ARA), trial 61 of 200, RNG seed 3082009981. Full reproduction assets are in reproduce/.
| File | Format | Size |
|---|---|---|
model-00001-of-00003.safetensors ... model-00003-of-00003.safetensors | BF16 | (see repo files) |
Yi-1.5-6B-Chat-heretic-ara-F16.gguf | GGUF F16 | 12.12 GB |
Yi-1.5-6B-Chat-heretic-ara-Q4_K_M.gguf | GGUF Q4_K_M | 3.67 GB |
Yi-1.5-6B-Chat-heretic-ara-Q5_K_M.gguf | GGUF Q5_K_M | 4.30 GB |
Yi-1.5-6B-Chat-heretic-ara-Q6_K.gguf | GGUF Q6_K | 4.97 GB |
Yi-1.5-6B-Chat-heretic-ara-Q8_0.gguf | GGUF Q8_0 | 6.44 GB |
GGUF quants are produced with llama.cpp. Run llama serve -hf saidutta69/Yi-1.5-6B-Chat-heretic-ara to pull the default quant.
bash
# llama.cppllama serve -hf saidutta69/Yi-1.5-6B-Chat-heretic-ara:Q4_K_Mpython
# transformersfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "saidutta69/Yi-1.5-6B-Chat-heretic-ara"tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype="auto", device_map="auto",) messages = [{"role": "user", "content": "Write a Python function to merge two sorted lists."}]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=512, 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.
Yi-1.5-6B-Chat-heretic-ara is a decensored build of 01-ai/Yi-1.5-6B-Chat produced with Heretic 2.0.0 (Arbitrary-Rank Ablation). It is the same weights with the refusal direction removed, not a new fine-tune, so its capabilities track the base model.
Yes. On Heretic's 100-prompt refusal harness it drops from 87/100 to 9/100 at KL divergence 0.0169 from the base model. These are measured values, not a claim; reproduction data is in reproduce/.
llama serve -hf saidutta69/Yi-1.5-6B-Chat-heretic-ara:Q4_K_M pulls the default quant; the repo ships 5 GGUF files (F16 plus Q4_K_M, Q5_K_M, Q6_K and Q8_0 — see Files).
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.
Inherits the Apache-2.0 license from the base model.
01-ai/Yi-1.5-6B-Chat - the base model this checkpoint derives fromRecorded in reproduce/reproduce.json:
Linux-6.6.0-hiveos-x86_64-with-glibc2.35Heretic notes that KL-divergence and refusal numbers can vary slightly with hardware and library versions.
Made with ❤️ by RACER IS OP (saidutta69) - follow for more uncensored models
# llama.cppllama serve -hf saidutta69/Yi-1.5-6B-Chat-heretic-ara:Q4_K_M# transformersfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "saidutta69/Yi-1.5-6B-Chat-heretic-ara"tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype="auto", device_map="auto",) messages = [{"role": "user", "content": "Write a Python function to merge two sorted lists."}]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=512, do_sample=False)print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))