TildeOpen-30b-64k

Developed by: Tilde.ai
Paper: https://arxiv.org/abs/2603.08182
Funded by: European Commission via EuroHPC JU Large AI Grand Challenge
Model type: A 30B parameter dense decoder-only transformer
Languages: Albanian, Bosnian, Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Hungarian, Icelandic, Irish, Italian, Latgalian, Latvian, Lithuanian, Macedonian, Maltese, Montenegrin, Norwegian, Polish, Portuguese, Romanian, Russian, Serbian, Slovak, Slovene, Spanish, Swedish, Turkish, Ukrainian as well as mathematical proofs, programming code and XML documents containing translation data
License: CC-BY-4.0

Info

This is the large context version of TildeOpen 30B foundational model, featuring context extension from 8k to 64k tokens using YaRN. We used a mixture of TildeOpen 30B base training data, HuggingFace's FinePDF datset and synthetic data based on KNOT. Due to YaRN, this model is primarily intended for use with transformers >= 5 or vLLM. We provide auto patches for use with older transformers versions, explained in this differences section.

For more detailed background information, please refer to the original model repository: https://huggingface.co/TildeAI/TildeOpen-30b.

This model remains a base model and has not been instruction tuned, nor human aligned.

Model Hyper-Parameters

We use the following YaRN configuration for RoPE scaling:

We follow Deespeek v3 and slightly overscale the YaRN embeddings to 10x rather than 8x; related PyTorch warnings can be ignored.

Running model using HF transformers >= 5

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

# Load tokenizer + model
tokenizer = AutoTokenizer.from_pretrained("TildeAI/TildeOpen-30b-64k")
model = AutoModelForCausalLM.from_pretrained(
    "TildeAI/TildeOpen-30b-64k",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Tokenize
inputs = tokenizer(user_in, return_tensors="pt").to(model.device)

# Generate (greedy, deterministic)
outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    repetition_penalty=1.2,
    do_sample=False,
)

text = tokenizer.decode(outputs[0], skip_special_tokens=True)

Running model using (old) HF transformers < 5

NOTE: The provided YARN patch was written specifically for transformers==4.46.3. It likely can support other versions, but that has not been thoroughly tested. We suggest avoiding patches and using transformers >= 5 or vLLM.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

# Load tokenizer + model
tokenizer = AutoTokenizer.from_pretrained("TildeAI/TildeOpen-30b-64k")
model = AutoModelForCausalLM.from_pretrained(
    "TildeAI/TildeOpen-30b-64k",
    trust_remote_code=True, # important, this applies patches 
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Tokenize
inputs = tokenizer(
    user_in,
    return_tensors="pt",
    return_token_type_ids=False, # sometimes needed for older transformers
).to(model.device)

# Generate (greedy, deterministic)
outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    repetition_penalty=1.2,
    do_sample=False,
)

text = tokenizer.decode(outputs[0], skip_special_tokens=True)

Evaluation

Needle-in-a-Haystack (NIAH) Evaluation for context extension

Smaller "Needle Position" values indicate placement closer to the beginning of the document; larger values indicate placement closer to the end.

Experiment Setup

For each context length and needle insertion depth:

1. The target context length was selected from Paul Graham’s essays.
2. The sentence The best thing to do in San Francisco is eat a sandwich and sit in Dolores Park on a sunny day. was inserted between two sentences around the target depth in the document.
3. The sentence The best thing to do in San Francisco is was appended.
4. The model output was considered correct if the model generated eat a sandwich and sit in Dolores Park on a sunny day within the next 15 tokens.

Alternative NIAH Experiment

Since the classic needle test has been known for a long time, there is risk that the model overfit to this exact sentence. So to double check the capabilities we also ran the same experiment, but with the string tripplet The secret key is 954834; The secret key is and 954834 and 5 different samples of text.

Belebele Benchmark: Reading Comprehension

What is Belebele Benchmark? Belebele is a multiple-choice machine reading comprehension (MRC) dataset spanning 122 language variants. This dataset enables the evaluation of mono- and multi-lingual models in high-, medium-, and low-resource languages. Each question has four multiple-choice answers and is linked to a short passage from the FLORES-200 dataset. The human annotation procedure was carefully curated to create questions that discriminate between different levels of generalizable language comprehension and is reinforced by extensive quality checks. Results

Why does this Matter? Belebele tests LLM's ability to provide answers based on a given text -- a standard use case in retrieval augumented generation workflows.

What did we do? We used the standard implementation of the belebele task from the LLM Evaluation Harness. We report 5-shot accuracy.

MultiBLiMP Benchmark: Grammar Test

What is MultiBLiMP? MultiBLiMP is a massively multilingual test of core grammar. It gives models pairs of almost-identical sentences—one grammatical and one ungrammatical—and asks whether the model assigns a higher probability to the correct one. Version 1.0 covers 101 languages

Why does this Matter? MultiBLiMP tests models' ability to distinguish correct and erroneous language. Just like humans, producing mostly correct language is not a big achievement. Rather, it is very bad to make any mistakes at all.

What did we do? We used the standard implementation of the MultiBLiMP task from the LLM Evaluation Harness. We report 0-shot accuracy.

Knowledge tests

ARC Benchmark Results

What is ARC? ARC - The AI2 Reasoning Challenge is a multiple-choice science question benchmark in English, derived from U.S. grade-school standardized exams. It has two subsets — ARC Easy and ARC Challenge — designed to test factual knowledge and common-sense.

Why does this Matter? ARC probes a model’s ability to answer non-trivial questions by applying world knowledge. Although the answer can sometimes be inferred from the question, in the classic lm-evaluation-harness ARC implementation the answer choices for each question are not provided during inference, thus placing emphasis on world knowledge, rather than on the model's reasoning capabilities.

What did we do? We use multilingual translations of ARC provided by Eurolingua; please refer to the publication. Other than the data source, we replicate the standard LM Evaluation Harness configuration for ARC. Our exact configuration is available at [TBA]. We report 5-shot accuracy.

MMLU Benchmark Results

What is MMLU? MMLU is a massive multitask test consisting of multiple-choice questions from various branches of knowledge, in English. The test spans subjects in the humanities, social sciences, hard sciences, and other areas that are important for some people to learn. This covers 57 tasks including elementary mathematics, US history, computer science, law, and more. To attain high accuracy on this test, models must possess extensive world knowledge and problem solving ability. Questions are four option multiple choice and assess factual knowledge, reading comprehension, and reasoning across disciplines. The questions can be grouped under four topics - stem, humanities, social_sciences and other, allowing for individual evaluation of each group.

Why does this Matter? Similarly to ARC, MMLU measures broad, general purpose factual knowledge and some reasoning capabilites. The possible answer choices are included during prompting, which can allow the model to employ reasoning to discard false answers, rather than just relying on knowing the correct one. It should be noted that some question groups are exclusive to the anglocentric world, e.g. US history or law.

What did we do? We use multilingual translations of MMLU provided by Eurolingua, please refer to the publication. Other than the data source, we replicate the standard LM Evaluation Harness configuration for MMLU. Our configuration is available at [TODO]. We report 0-shot accuracy.

National Exams Results

What are National Exams? A curated suite of multlingual publicly available past questions from national-level standardized exams across multiple countries (e.g., high-school exit and university-entrance exams), please refer to the publication. The dataset is available on HuggingFace here. Items are presented in multiple-choice format.

Why does this Matter? Similarly to MMLU, the model is tested on factual knowledge and reasoning capabilites. However, it should be stressed that for each language the bench is unique (the exams are different) and available in the source language (i.e. not translated). This places emphasis on the model's regional knowledge and eliminates translation noise that is present in many other multilingual benchmarks. Possible answer choices are once again included during inference, allowing for the model to employ reasoning if factual knowledge is lacking.

What did we do? [TODO]

Differences from Huggingface LLaMa model implementation for transformers <5

Patch is primarily contained in llama_yarn_patch_4x.py and configuration_llama_patch_4x.py. We provide wrappers for enabling AutoModel API. Running patched code with transformers >= 5 is safe and will ignore the patch files and default to base implementation.

Main differences:

• New rotary embedding class was written - NeoXRotaryEmbeddings.

  • Supports YaRN, implemented by analogy with vLLM’s YaRN approach and the YARN paper
  • Designed to match the rotary embedding behaviour used during pretraining and context extension.

• The attention implementation was modified to more closely match the attention behaviour used during training.

Running the code requires flash-attn >= 2.0.6, < 3.0. The model can still be run with the original Hugging Face LLaMa code. However, when using YaRN, we found that can lead to vastly different logit generation.

NOTE: If you are using transformers >= 5 or vLLM, this section does not apply and can safely be ignored.