GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
yisberh
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
yisberh
Model Tree
Input Modalities
Output Modalities
Supported Functionality
Fidel Law 1 should not be described as:
“A 23M model fine-tuned from our base pretrained model.”
That description would incorrectly suggest that Fideliz AI started with an external foundation model.
Instead:
Fidel Law 1 is a 23M-parameter legal language model developed on Fideliz AI's own internally developed base model.
The legal model was built using Fideliz AI's own model-development pipeline and legal training data.
The distinction is important because building from scratch is a core part of Fideliz AI's research mission.
Our objective is to develop the underlying AI technology ourselves rather than simply adapting an existing foreign foundation model.
The Fidel Law 1 development process covers the full model-development pipeline.
text
Ethiopian Legal Sources ↓Data Collection ↓Data Cleaning & Preparation ↓Legal Dataset Construction ↓Question & Reasoning Data ↓Data Validation & Evaluation ↓Amharic Tokenizer ↓Fideliz AI Base Model ↓Legal Supervised Training ↓Fidel Law 1 ↓Evaluation & Error Analysis ↓Model ImprovementThe purpose of this pipeline is to create an AI model that understands Ethiopian legal language rather than simply reproducing generic language-model behavior.
Fidel Law 1 is designed for:
The model performs best with short, direct legal questions in Amharic.
Potential applications include:
Fidel Law 1 should not be used as the sole basis for:
| Property | Value |
|---|---|
| Parameters | ~22M |
| Architecture | LLaMA-style decoder-only Transformer |
| Layers | 8 |
| Hidden size | 512 |
| Attention heads | 8 |
| Key/Value heads | 4 |
| Vocabulary size | 8,000 |
| Maximum context | 100 tokens |
| Language | Amharic |
The legal training data focuses primarily on Ethiopian legislation and legal reasoning.
Major areas include:
The project emphasizes locally relevant Ethiopian legal data rather than relying exclusively on generic multilingual legal datasets.
Fidel Law 1 was developed using Fideliz AI's internally developed base model and training pipeline.
The training process involved multiple stages, including:
The legal training stage teaches the internally developed base model to specialize in Ethiopian legal language and legal question answering.
Legal text was transformed into structured training examples covering:
The training data was designed to encourage the model to identify and reference relevant legal provisions.
Fidel Law 1 was evaluated using internally constructed Ethiopian legal evaluation datasets.
| Evaluation | Result |
|---|---|
| General Legal Retrieval | 96% |
| Criminal Code Case-to-Article | **95.50% ** |
| Bidirectional Article Mapping | **100% ** |
| Direct Article Identification, Arts. 1–865 | **86.67% ** |
| Complex Case-to-Article Reasoning | **85% ** |
These are internal empirical evaluations and are not standardized external benchmarks.
Tests whether the model can identify the relevant legal provision from a natural-language legal question.
Tests whether the model can map a described legal situation to the relevant Ethiopian Criminal Code article.
Tests both:
text
Article → Legal ConceptLegal Concept → ArticleTests whether the model can identify specific legal provisions when directly queried.
Tests more difficult legal scenarios requiring the model to connect facts with the relevant legal provision.
The following examples demonstrate Fidel Law 1's Amharic legal question-answering behavior.
🗣️ ጥያቄ 1:
አንቀጽ 1 ዓላማው ምንድነው?
🤖 መልስ:
የወንጀል ሕግ ዓላማ የአገሪቱን መንግሥት፣ የሕዝቦቹን፣ የነዋሪዎቹን፣ ሰላም፣ ደህንነት፣ ሥርዓት፣ መብትና ጥቅም መጠበቅና ማረጋገጥ ነው። [SOURCE: 1]
🗣️ ጥያቄ 2:
የወንጀል ሕግ ግብ ምንድን ነው?
🤖 መልስ:
ወንጀል እንዳይፈጸም መከላከል ነው [ARTICLE: 1]።
🗣️ ጥያቄ 3:
ፍርድ ቤት ያልተደነገገ ድርጊትን ወንጀል ማድረግ ይችላል?
🤖 መልስ:
አይችልም። 2(2) መሠረት ሕገወጥነቱ በሕግ ያልተደነገገ ድርጊት ወይም ግድፈት እንደወንጀል ሊቆጠርና ቅጣት ሊወሰንበት አይችልም። [SOURCE: 2]
🗣️ ጥያቄ 4:
ሕጉ አሻሚ ከሆነ ምን ይደረጋል?
🤖 መልስ:
2(4) መሠረት ፍርድ ቤቱ በሕግ አውጪው ባቀደው ትርጉምና በሕጉ መንፈስ መሠረት መተርጐም አለበት። [SOURCE: 2]
🗣️ ጥያቄ 5:
አንቀጽ 4 ለማን ይሠራል?
🤖 መልስ:
አንቀጽ 4 የሕግ ፊት እኩልነትን ይመለከታል። [SOURCE: 4]
🗣️ ጥያቄ 6:
በጾታ ልዩነት ይደረጋል?
🤖 መልስ:
የወንጀል ሕግ የጾታን ልዩነት ሳያደርግ በሁሉም ላይ በእኩልነት ተፈፃሚ ይሆናል። [SOURCE: 4]
🗣️ ጥያቄ 7:
አዲሱ ሕግ ቀድሞ ለተፈፀመ ድርጊት ይሠራል?
🤖 መልስ:
አንቀጽ 5 የወንጀል ሕግ ወደኋላ ተመልሶ የማይሠራ መሆኑን ይደነግጋል። [SOURCE: 5]
🗣️ ጥያቄ 8:
6 ላይ የተሻለ ሕግ ማን ይወስናል?
🤖 መልስ:
ፍርድ ቤቱ ይወስናል። [SOURCE: 6]
🗣️ ጥያቄ 9:
7 የትኛው ድንጋጌ ይተገበራል?
🤖 መልስ:
የዚህ ሕግ ድንጋጌዎች ከአንቀጽ 129 እስከ 185 ተፈፃሚ ይሆናሉ። [SOURCE: 7]
🗣️ ጥያቄ 10:
የግል አቤቱታ መብት መቼ ቀሪ ይሆናል?
🤖 መልስ:
ይህ ሕግ ከፀናበት ጊዜ ጀምሮ በሁለት ዓመት ውስጥ አቤቱታ አቅራቢው መብቱን ካልተጠቀመበት ቀሪ ይሆናል። [SOURCE: 8(2)]
Fidel Law 1 has a maximum context length of 100 tokens.
The current version is therefore optimized for:
It is not designed for long legal documents or extended conversations.
Fidel Law 1 is designed to produce concise legal answers.
When possible, responses provide:
Example:
text
የወንጀል ሕግ ወደኋላ ተመልሶ አይሠራም። [SOURCE: 5]Known limitations include:
The model should therefore be treated as an experimental Ethiopian legal AI research model.
Fidel Law 1 is an experimental research model and is not a source of binding legal advice.
Model outputs can contain errors, omissions, incorrect interpretations, or incorrect article references.
Always verify legal information against authoritative Ethiopian legal sources and applicable legislation before relying on an output for legal, professional, governmental, judicial, or other consequential purposes.
python
from transformers import AutoTokenizer, AutoModelForCausalLMimport torch repo_id = "yisberh/fidel_law_1"device = torch.device("cuda" if torch.cuda.is_available() else "cpu") tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)model = AutoModelForCausalLM.from_pretrained(repo_id).to(device)model.eval() prompt = ( "<|im_start|>user\n" "አንቀጽ 1 ዓላማው ምንድነው?" "<|im_end|>\n" "<|im_start|>assistant\n") # ==============================================================================# 🛠️ THE FIX: BYPASS THE BUGGY `tokenizer()` WRAPPER# ==============================================================================bos_id = tokenizer.bos_token_ideos_id = tokenizer.eos_token_idim_end_id = tokenizer.convert_tokens_to_ids("<|im_end|>") blocked_tokens = [ tokenizer.convert_tokens_to_ids("<|im_start|>"), tokenizer.convert_tokens_to_ids("<|system|>"), tokenizer.convert_tokens_to_ids("<|user|>"), tokenizer.convert_tokens_to_ids("<|assistant|>")] # 1. Encode manually and prepend the BOS tokeninput_ids = [bos_id] + tokenizer.encode(prompt, add_special_tokens=False) # 2. Build the PyTorch tensors manuallyinput_tensor = torch.tensor([input_ids], dtype=torch.long, device=device)attention_mask = torch.ones_like(input_tensor, device=device) with torch.no_grad(): outputs = model.generate( input_ids=input_tensor, attention_mask=attention_mask, max_new_tokens=80, pad_token_id=eos_id, eos_token_id=[eos_id, im_end_id], do_sample=False, # Pure argmax for factual legal text suppress_tokens=blocked_tokens # Prevents structural token bleeding ) # Extract only the assistant's newly generated tokensinput_length = input_tensor.shape[1]generated_ids = outputs[0][input_length:] response = tokenizer.decode(generated_ids).strip().replace("<|im_end|>", "")print(response)For deterministic legal retrieval and evaluation:
python
outputs = model.generate( **inputs, max_new_tokens=80, do_sample=False)Deterministic generation can make evaluation and reproducibility easier.
Detailed carbon-emission measurements are not currently available for this release.
Fidel Law 1 uses a decoder-only Transformer architecture developed within Fideliz AI's model-development program.
The model contains:
The legal model was developed on top of the Fideliz AI internally developed base model.
GPU acceleration was used throughout model development and training.
Fidel Law 1 represents a broader research direction at Fideliz AI:
Build AI, don't just adapt AI.
Our objective is to develop the underlying model capabilities, training processes, datasets, evaluation systems, and domain-specific AI systems ourselves.
For Ethiopian AI, this means developing models that can understand Ethiopian languages, legal systems, culture, knowledge, and real-world contexts.
The long-term goal is to build increasingly capable African foundation models from the ground up.
If you use Fidel Law 1 in research, please cite:
bibtex
@misc{bule2026fidellaw1, author = {Yisak Bule and Fideliz AI}, title = {Fidel Law 1: An Amharic Legal Language Model for Ethiopian Legal Text}, year = {2026}, howpublished = {Hugging Face}, url = {https://huggingface.co/yisberh/fidel_law_1}}Bule, Y., & Fideliz AI. (2026). Fidel Law 1: An Amharic Legal Language Model for Ethiopian Legal Text. Hugging Face.
Fidel Law 1
An Amharic legal language model developed by Fideliz AI.
Fideliz AI Base Model
The internally developed base model on which Fidel Law 1 was developed.
Legal Retrieval
Identifying relevant legal provisions from a natural-language query.
Case-to-Article Mapping
Mapping a described legal situation to a relevant legal article.
Bidirectional Mapping
Mapping an article to a legal concept and a legal concept back to its article.
Causal Language Model
A language model that predicts subsequent tokens based on previous tokens.
Custom Tokenizer
A tokenizer developed for the Amharic/Ge'ez language domain.
Fidel Law 1 is part of Fideliz AI's broader effort to develop reasoning-focused AI and foundation models from the ground up in Ethiopia.
The company's work as building foundation AI models from zero, with research spanning mathematics, biology, law, finance, and Ethiopian applications.
Fideliz AI is also actively developing Ethiopian training data, evaluation processes, and locally relevant AI systems.
Fideliz AI:
https://www.linkedin.com/company/fideliz-ai
Yisak Bule
AI Team Lead, Fideliz AI
Fideliz AI 🇪🇹
For research collaboration, technical questions, or feedback, please contact the Fideliz AI team through the project repository or company page.
Fidel Law 1 is released under the Apache License 2.0.
See the repository license for the complete terms.
Ethiopian Legal Sources ↓Data Collection ↓Data Cleaning & Preparation ↓Legal Dataset Construction ↓Question & Reasoning Data ↓Data Validation & Evaluation ↓Amharic Tokenizer ↓Fideliz AI Base Model ↓Legal Supervised Training ↓Fidel Law 1 ↓Evaluation & Error Analysis ↓Model ImprovementArticle → Legal ConceptLegal Concept → Articleየወንጀል ሕግ ወደኋላ ተመልሶ አይሠራም። [SOURCE: 5]from transformers import AutoTokenizer, AutoModelForCausalLMimport torch repo_id = "yisberh/fidel_law_1"device = torch.device("cuda" if torch.cuda.is_available() else "cpu") tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)model = AutoModelForCausalLM.from_pretrained(repo_id).to(device)model.eval() prompt = ( "<|im_start|>user\n" "አንቀጽ 1 ዓላማው ምንድነው?" "<|im_end|>\n" "<|im_start|>assistant\n") # ==============================================================================# 🛠️ THE FIX: BYPASS THE BUGGY `tokenizer()` WRAPPER# ==============================================================================bos_id = tokenizer.bos_token_ideos_id = tokenizer.eos_token_idim_end_id = tokenizer.convert_tokens_to_ids("<|im_end|>") blocked_tokens = [ tokenizer.convert_tokens_to_ids("<|im_start|>"), tokenizer.convert_tokens_to_ids("<|system|>"), tokenizer.convert_tokens_to_ids("<|user|>"), tokenizer.convert_tokens_to_ids("<|assistant|>")] # 1. Encode manually and prepend the BOS tokeninput_ids = [bos_id] + tokenizer.encode(prompt, add_special_tokens=False) # 2. Build the PyTorch tensors manuallyinput_tensor = torch.tensor([input_ids], dtype=torch.long, device=device)attention_mask = torch.ones_like(input_tensor, device=device) with torch.no_grad(): outputs = model.generate( input_ids=input_tensor, attention_mask=attention_mask, max_new_tokens=80, pad_token_id=eos_id, eos_token_id=[eos_id, im_end_id], do_sample=False, # Pure argmax for factual legal text suppress_tokens=blocked_tokens # Prevents structural token bleeding ) # Extract only the assistant's newly generated tokensinput_length = input_tensor.shape[1]generated_ids = outputs[0][input_length:] response = tokenizer.decode(generated_ids).strip().replace("<|im_end|>", "")print(response)outputs = model.generate( **inputs, max_new_tokens=80, do_sample=False)@misc{bule2026fidellaw1, author = {Yisak Bule and Fideliz AI}, title = {Fidel Law 1: An Amharic Legal Language Model for Ethiopian Legal Text}, year = {2026}, howpublished = {Hugging Face}, url = {https://huggingface.co/yisberh/fidel_law_1}}| Objective | Causal Language Modeling |
| Base model | Fideliz AI internally developed base model |
| Tokenizer | Custom Ge'ez/Amharic BPE |