This is a decensored version of llm-jp/llm-jp-4.1-33b-thinking, made using custom fork of Heretic
Abliteration parameters
| Parameter | Value |
|---|---|
| target_layers | 34, 36, 38, 40–42, 45–47, 49 |
| layer_weights | ×2 on layers 34, 36, 38, 40, 42, 46; ×1 elsewhere |
| preserve_good_behavior_weight | 1.0 |
| steer_bad_behavior_weight | 0.5 |
| overcorrect_relative_weight | 0.7 |
| neighbor_count | 1 |
| ridge_regularization | 0.0024 |
| transport_rank | 4 |
| entropy_regularization | 0.1 |
| transport | gaussian |
| lora_rank | 128 |
| row_normalization | none |
| target_components | attn.o_proj, mlp.down_proj |
| covariance_regularization | 0.01 |
| max_weight_change | 1.0 |
Performance
| Metric | This model | Original model (llm-jp/llm-jp-4.1-33b-thinking) |
|---|---|---|
| Refusals | 0/100 | 88/100 |
| KL divergence | 0.0118 | 0 (by definition) |
| Benchmark | Original model | This model |
|---|---|---|
| MMLU (all 14,042, 0-shot, lm-eval prompt) | 74.37% | 74.05% (−0.32, p = 0.051) |
| HumanEval pass@1 (164) | 94.51% | 93.29% (−1.22, p = 0.69) |
Measured on these weights quantized to Q4_K_M, against the official Q4_K_M. Refusals: keyword matching on the first 100 tokens of the reasoning (reasoning_effort low) for 100 Japanese harmful prompts. KL divergence: first-token KL on 100 Japanese prompts (OS-Software/harmless_alpaca_ja, test[:100]).
⚠️ Important Notice
This model has undergone substantial reduction of its safety alignment. As a result, it is more likely than standard models to generate harmful, inaccurate, biased, offensive, or otherwise inappropriate content.
Intended Use
For research and experimentation only, including safety research, alignment studies, and red-teaming. Please avoid deploying it in public or end-user-facing services.
User Responsibility
All outputs should be treated as untrusted and independently verified before use. Users are solely responsible for:
- Evaluating the accuracy and suitability of generated content
- Implementing appropriate safeguards and human oversight
- Complying with applicable laws, regulations, licenses, and ethical standards
Use of this model is entirely at your own risk.
Disclaimer
OS-Software provides this model without warranties of any kind and assumes no liability for any direct or indirect damages, losses, misuse, or legal consequences arising from its use.
Acknowledgements
Thanks to the base model developers, p-e-w for Heretic, and the wider open-source community.
This is a derivative work released under the base model’s applicable license. All rights to the base model remain with their respective owners.
llm-jp-4.1-33b-thinking
LLM-jp-4.1 is a series of large language models developed by the Research and Development Center for Large Language Models at the National Institute of Informatics.
This repository provides the llm-jp-4.1-33b-thinking model. For an overview of the LLM-jp-4.1 models across different parameter sizes, please refer to:
Base models are trained with pre-training and mid-training only. Post-trained models are aligned using supervised fine-tuning (SFT) and direct preference optimization (DPO), without reinforcement learning.
For more details on the training procedures and evaluation results, please refer to our technical blog (in Japanese).
For practical usage examples and detailed instructions on how to use the models, please also refer to our cookbook.
To support the continued development of LLM-jp, we would greatly appreciate it if you could share how you utilize LLM-jp outcomes via the survey form.
Usage
Please refer to our cookbook for practical usage examples and detailed instructions on how to use the models.
Model Details
- Model type: Transformer-based Language Model
- Architectures:
Dense model:
| Params | Layers | Hidden size | Heads | Context length | Embedding parameters | Non-embedding parameters | Total parameters |
|---|---|---|---|---|---|---|---|
| 8B | 32 | 4,096 | 32 | 65,536 | 805,306,368 | 7,784,894,464 | 8,590,200,832 |
| 33B | 64 | 5,120 | 40 | 65,536 | 1,006,632,960 | 32,212,915,200 | 33,219,548,160 |
MoE model:
| Params | Layers | Hidden size | Heads | Routed Experts | Activated Experts | Context length | Embedding parameters | Non-embedding parameters | Activated parameters | Total parameters |
|---|---|---|---|---|---|---|---|---|---|---|
| 32B-A3B | 32 | 2,560 | 40 | 128 | 8 | 65,536 | 503,316,480 | 31,635,712,512 | 3,827,476,992 | 32,139,028,992 |
Tokenizer
The tokenizer of this model is based on a Unigram byte-fallback model implemented with huggingface/tokenizers.
The vocabulary entries were converted from llm-jp-tokenizer v4.0.
Please refer to README.md of llm-jp-tokenizer for details on the vocabulary construction procedure (the pure SentencePiece training does not reproduce our vocabulary).
[!NOTE] The chat template of this model is designed to be compatible with the OpenAI Harmony response format. However, the tokenizer differs from the one assumed by the
openai-harmonylibrary, and therefore direct tokenization withopenai-harmonyis not supported. For correct behavior, please use the tokenizer provided with this model. For detailed usage, please refer to our cookbook.
Training
Pre-training
This model was trained through a multi-stage pipeline consisting of pre-training and mid-training phases, using a total of 11.7T tokens.
The corpora used for pre-training and mid-training are publicly available at the following links:
[!NOTE] Although most of the corpora have been released, some portions are excluded from public release due to licensing constraints.
Post-training
We have fine-tuned the pre-trained checkpoint using SFT and further aligned it with DPO.
The datasets used for post-training are also publicly available at the following links:
- SFT
- DPO (for llm-jp-4.1-8b-thinking model)
- DPO (for llm-jp-4.1-32b-a3b-thinking model)
- DPO (for llm-jp-4.1-33b-thinking model)
Evaluation
We evaluated llm-jp-4.1 on a variety of benchmarks covering general capabilities, safety, and tool calling.
For more detailed evaluation results and analysis, please refer to our technical blog.
swallow-evaluation-instruct
We evaluated the models on a range of benchmarks covering the following six categories:
- Math
- Math 500
- AIME 2024 (pass@1, pass@32)
- AIME 2025 (pass@1, pass@32)
- AIME 2026 (pass@1, pass@32)
- MCLM Math 100 (pass@1, pass@4)
- PolyMath JA High
- PolyMath JA Top
- Science
- GPQA Diamond (pass@1, pass@4)
- JGPQA Diamond
- Knowledge & QA
- JAM-CQA
- JEMHopQA
- JMMLU
- MMLU-ProX JA
- MMLU-ProX EN
- Code
- LiveCodeBench v6 (pass@1, pass@10)
- JHumanEval (pass@1, pass@10)
- HumanEval+ (pass@1, pass@10)
- Instruction Following (IF)
- MIFEval JA
- IFBench
- Machine Translation (MT)
- WMT20 EN-JA
- WMT20 JA-EN
For LLM-jp and gpt-oss models, reasoning_effort was set to high.
For Olmo-3-7B-Think, Olmo-3.1-32B-Think, Qwen3, Qwen3.5, Qwen3.6, and Gemma 4, enable_thinking was set to True.
For Qwen3.8-27B and Muse-Glimmer-30B, reasoning_effort was set to xhigh.
The figure below shows the average score across the benchmarks in each category.
llm-jp-judge
We evaluated the models using an LLM-as-a-Judge framework on the following benchmarks:
- MT-Bench (JA/EN): A benchmark for measuring multi-turn conversational task-solving ability.
- AnswerCarefully: A benchmark for evaluating safety in Japanese. We used 336 questions from the v2.0 test set.
- llm-jp-instructions: A set of human-created single-turn question-answer pairs. We used 400 questions from the test set.
We used gpt-5.4-2026-03-05 as the judge. For models that support reasoning_effort, it was set to medium.
The scores represent the average values obtained from three rounds of inference and evaluation. For more details, please refer to the evaluation code.
| Model Name | MT-Bench (JA) | MT-Bench (EN) | AnswerCarefully | llm-jp-instructions |
|---|---|---|---|---|
| gpt-4o-2024-08-06 | 7.29 | 7.69 | 4.00 | 4.07 |
| gpt-5.4-2026-03-05 | 8.87 | 8.89 | 4.43 | 4.82 |
| gpt-oss-20b | 7.33 | 7.85 | 3.55 | 3.16 |
| llm-jp-4-8b-thinking | 7.54 | 7.79 | 3.69 | 3.54 |
| llm-jp-4.1-8b-thinking | 7.58 | 7.67 | 3.92 | 3.67 |
| llm-jp-4-32b-a3b-thinking | 7.82 | 7.86 | 3.70 | 3.61 |
| llm-jp-4.1-32b-a3b-thinking | 7.69 | 7.85 | 3.91 | 3.79 |
| llm-jp-4-33b-thinking | 8.00 | 8.24 | 3.79 | 3.79 |
| llm-jp-4.1-33b-thinking | 7.76 | 7.98 | 4.08 | 3.83 |
Tool Calling
We evaluated the models on the following tool calling benchmarks:
For BFCL, we evaluated the models using only the categories available up to v3, so the web search and memory categories introduced in v4 are excluded.
For tau2-bench, we used Azure's gpt-5.1-2025-11-13 as both the user simulator and the NL-assertion judge, and the scores represent the average values obtained from two rounds of inference and evaluation.
For all LLM-jp models, reasoning_effort was set to medium.
The figure below shows the scores on each benchmark.
Risks and Limitations
The models released here are research and development models and are not intended for direct use in production services. Although the models have undergone post-training for instruction following and safety, they may still generate inaccurate, inappropriate, or otherwise undesirable outputs. Users should carefully evaluate the models for their intended use cases.
Send Questions to
llm-jp(at)nii.ac.jp
License
Acknowledgements
To develop this model, we used the NINJAL Web Japanese Corpus (whole-NWJC) from the National Institute for Japanese Language and Linguistics (NINJAL).
Model Card Authors
The names are listed in alphabetical order.
Hirokazu Kiyomaru, Takashi Kodama, and Yunang Wu.