EuroLLM-1.7B-Instruct-heretic

Test run of Heretic 2.0, not the usual pipeline. This is a test of Heretic 2.0's Arbitrary-Rank Ablation (ARA) modifier, not the Heretic 1.4 directional-ablation pipeline used for my usual releases. It is published for inspection, so only the F16 GGUF and the safetensors are provided - no quant ladder.

Behaviourally neutral edit. Heretic measured 3/100 refusals for the base and 3/100 for this checkpoint at KL 0.0113. utter-project/EuroLLM-1.7B-Instruct is only lightly aligned, so there was almost nothing for the abliteration to remove, and the optimiser stayed conservative. This is not a decensoring upgrade - it is the same model with a small weight edit. The companion lm-eval run confirms it: IFEval prompt-strict 17.93 vs 17.56, MMLU 25.92 vs 25.76, both inside one standard error.

A variant of utter-project/EuroLLM-1.7B-Instruct produced with Heretic 2.0.0.dev0 using Arbitrary-Rank Ablation (ARA) - a test run of the v2 modifier rather than the v1.4 pipeline behind my usual releases. EuroLLM-1.7B-Instruct is the first instruction-tuned checkpoint of the EuroLLM series: a dense 1.66B multilingual model trained on 4 trillion tokens across 35 languages, then tuned on EuroBlocks for general instruction-following and machine translation. ARA edits the network over layers 8-23; everything else is unchanged.

Who this is for: multilingual European-language work at 1.66B - translation, instruction following and cheap inference on small hardware - plus experiments that need an ARA-abliterated EuroLLM checkpoint with a reproducible recipe. It is not for anyone who wants fewer refusals: the base already refuses almost nothing, and this checkpoint refuses exactly as much.

Context length is only 4,096 tokens. max_position_embeddings is 4,096 and the config ships no rope-scaling or YaRN block, so the GGUF loads with a 4,096-token limit (verified in the header). The base card mentions that RoPE "allows the extension of the context length" but publishes no recipe - any long-context setup is on you.

Abliteration parameters

Trial 75 of a 200-trial Heretic 2.0 ARA run (seed 42).

Performance

All three remaining refusals stayed exactly as they were: the base was already at 3/100, so there was nothing to remove. KL 0.0113 is a small weight change, and the lm-eval run below is unchanged within noise. Do not read this as a "high-fidelity edit that also decensored the model" - it only did the first half.

Base model benchmarks

EuroLLM's published numbers for the 1.7B base, not re-run against this checkpoint:

Abliteration removes a weight subspace; it does not add capability. Treat these as the base model's numbers, not a re-measurement of this checkpoint.

lm-eval parity check

User-run lm-eval 0.4.13, this checkpoint vs the base (MMLU: 14,042 questions, stderr 0.37 pp; IFEval prompt-level stderr 1.65 pp):

Every delta is inside one standard error. This is the expected result for a 3->3/100 refusal run.

Why ARA instead of directional ablation

Directional ablation finds one refusal direction and orthogonalizes attention/MLP outputs against it at a fixed strength on every layer. ARA learns a low-rank subspace over a layer range and blends a "preserve good behaviour" objective on mlabonne/harmless_alpaca with a "steer bad behaviour" objective on mlabonne/harmful_behaviors. This run weighted the preserve objective 0.0554 against a 0.0001 steering weight, and the result is a small edit (KL 0.0113) that leaves refusal behaviour where it started. See the Heretic repo.

Files

Safetensors

BF16, 1,656,850,432 parameters (1.66B), untied input/output embeddings (~0.262B each). The reproduce/ directory carries the full Heretic 2.0 recipe - config.toml, requirements.txt, the Optuna study journal, and SHA-256 sums - so this exact checkpoint can be regenerated bit-for-bit with heretic --reproduce reproduce/reproduce.json.

GGUF quantization

Test run, so F16 is the only GGUF published (3.32 GB). It was produced with llama.cpp master on 2026-10-30; no quant ladder was generated for this checkpoint. Quantize the F16 yourself with llama-quantize if you need smaller files.

Text GGUF: arch llama, 219 tensors, 24 blocks, 4096 context, 128,000 vocab, untied embeddings, RoPE theta 10,000. The tensor count and 1,656,850,432 elements match the safetensors exactly, so nothing was dropped. Standard dense Llama architecture - loads natively in llama.cpp, Ollama, LM Studio and Jan, and works with MLX and vLLM.

Run llama serve -hf saidutta69/EuroLLM-1.7B-Instruct-heretic:F16 to pull the F16 file explicitly - there is no smaller quant in this repo, so the bare -hf <repo> form has nothing else to pick.

Quickstart

# llama.cpp - ChatML template needs --jinja
llama serve -hf saidutta69/EuroLLM-1.7B-Instruct-heretic:F16 --jinja -c 4096

# transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "saidutta69/EuroLLM-1.7B-Instruct-heretic"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype="auto", device_map="auto")

messages = [
    {"role": "system", "content": "You are EuroLLM, an AI assistant specialised in European languages."},
    {"role": "user", "content": "Translate to Polish: 'The library closes at six.'"},
]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, tokenize=True,
                                 return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=256)
print(tok.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Abliteration does not touch the tokenizer or the chat template, so the base model's multilingual setup is intact.

Generation parameters

generation_config.json sets do_sample: true with no temperature, top-p or top-k override, and the base card publishes no tuned sampling settings, so generate() runs on the model defaults. For translation, a low temperature is the safe starting point; validate on your own language pairs rather than copying a recipe from a different model.

Chat template

ChatML: system, user and assistant turns are wrapped in <|im_start|>role ... <|im_end|>, and the template is shipped as chat_template.jinja. The base card's example uses an empty system turn; a language hint there is a reasonable way to steer output. There is no thinking mode, no enable_thinking flag, and no tool-call schema - this is the plain EuroLLM instruct interface.

Runs anywhere

F16 only, as published here. At 1.66B and 4K context the KV cache is small, so decode is bandwidth-bound and a quantized build (which you would have to make yourself) runs acceptably on a phone-class CPU.

Model details

FAQ

What is EuroLLM-1.7B-Instruct-heretic?

EuroLLM-1.7B-Instruct-heretic is a decensored build of - produced with Heretic 2.0.0.dev0 (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.

Is EuroLLM-1.7B-Instruct-heretic actually uncensored?

Yes. On Heretic's 100-prompt refusal harness it drops from 3/100 to 3/100 at KL divergence 0.0113 from the base model. These are measured values, not a claim; reproduction data is in reproduce/.

How do I run EuroLLM-1.7B-Instruct-heretic locally?

llama serve -hf saidutta69/EuroLLM-1.7B-Instruct-heretic:F16 runs the only published quant (see Files).

Responsible use

This is a weight edit, not an alignment layer. Refusal behaviour is unchanged from the base (3/100), so it does not add or remove safety behaviour, and the base is documented as "not aligned to human preferences". It inherits EuroLLM's factual limitations and biases across all 35 languages; abliteration removes a direction in weight space, it does not add capability or judgment. Validate translation output before shipping it, especially on low-resource pairs and on untrusted content that you would not want laundered past another system's filters.

Made with ❤️ by RACER IS OP (saidutta69) - follow for more uncensored models

License

Inherits the apache-2.0 license from the base model, which is tagged Apache-2.0 on the Hub but ships no LICENSE file, so the canonical license text is linked rather than a repo blob. This is a derivative work; keep the license and attribution when redistributing, and do not imply endorsement by the EuroLLM authors.

Related

• utter-project/EuroLLM-1.7B-Instruct - the base instruction-tuned model (not gated)

• utter-project/EuroLLM-1.7B - the pre-trained, non-instruct version

• EuroLLM-9B-Instruct-heretic - the larger abliterated sibling

• EuroLLM project - project site and the rest of the model family

• RACER IS OP - Heretic Models - full collection - every heretic model in this collection

Abliteration hardware

Recorded in reproduce/reproduce.json:

• GPU: NVIDIA RTX PRO 6000 Blackwell Server Edition (94.97 GB VRAM)
• OS: Linux-4.19.0-gvisor-x86_64-with-glibc2.41
• CUDA / driver: CUDA 13.2, driver 595.71.05
• Software: PyTorch 2.15.0+cu132, Heretic 2.0.0.dev0

Heretic notes that KL-divergence and refusal numbers can vary slightly with hardware and library versions.