Alpha
Contact
Implement · Boldt-DC

Can you run it?

Ownership levelPartialnone·limited·partial·substantial·fullAnalytical input D ยท 45.2/100

This page is a projection of the one entry record, the Reliability and Data control factors that Implement covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.

Which domain expands which factor
  • AssessUse & modify + Transparency
  • ImplementData control + Reliability
  • UseReliability
  • SupportTransparency

Install & run

Load the weights with transformers and use plain text completion (ev-card-1b, ev-card-350m).

Call AutoTokenizer.from_pretrained and AutoModelForCausalLM.from_pretrained with Boldt/Boldt-DC-1B or Boldt/Boldt-DC-350M. Then call generate on a German prompt such as "Berlin ist eine Stadt, wo".

The Boldt-DC-1B card's own sample loads Boldt/Boldt-1B, a different model trained with news data and a 4,096 context. Use the repo name above. Always load the tokenizer from the same repo.

Hardware & VRAM requirements

The bf16 weights are 0.94 GB for the 350M model (468M parameters) and 2.41 GB for the 1B model (1.205B parameters). With a 2,048-token context the KV cache is small, so both run on a laptop GPU or CPU. This is OneHill's inference from file sizes. The publisher gives no VRAM, latency or throughput figure. The paper names training hardware (8xA100 and 4xH100) and no GPU-hours (ev-hf-platform, ev-paper).

Serving stacks

The cards document transformers only. The config is a standard LlamaForCausalLM, which common runtimes read, but the publisher makes no statement for vLLM, llama.cpp or Ollama and nobody has tested them for this entry. No inference provider hosts the models (ev-hf-platform).

Safe-deployment controls & Deployment Ceiling

Deployment Ceiling: T0 (sandbox and research), conditional. These are base models with no safety tuning and no safety evaluation, so keep them behind a human reviewer.

  • Run your own toxicity, bias and factuality tests before any wider use.
  • Add input and output guardrails; no guard model or acceptable-use policy ships.
  • Add the Apache-2.0 text and a notice when you redistribute, since no LICENSE file ships.
  • Pin the repo revision and check the LFS SHA256; nothing is signed.
  • Confirm which repo your code loads (see Install).

Fine-tuning & adaptation

The paper fine-tunes both base models on the German SmolTalk2 subset with TRL: batch size 32, peak learning rate 5e-5, 3% warmup, loss on completions only, one H100 (ev-paper). Third parties have published German chat and embedding fine-tunes of the DC models (ev-hf-platform). No configs or fine-tuning guide ship, and the Boldt-1B-IT-Preview sibling was tuned on a different base and data mix (ev-siblings).

How this scores

The ownership factors this domain covers, drawn from the one entry record.

3

ReliabilityIs it reliable and good enough for the job?

Weak

Safety is 1 (no safety evaluation, stated by the publisher) and operational readiness is 2 (transformers sample only, no quantised builds or hardware guidance); these are base models, not deployable assistants (ev-card-1b, ev-paper).

How this scores (AOI sub-dimensions)
Operational2/5how practical it is to run, serve and maintain in productionThe weights are standard safetensors with a plain Llama architecture and load through transformers (the cards show only that route).
Safety1/5whether misuse risks are evaluated and guardrails are providedThese are base models and say so; the cards and the paper state plainly that no toxicity, bias or stereotype evaluation was done and that repeated epochs may amplify biases.
4

Doesn't extract your dataDoes running it keep your knowledge and data yours?

Strong

Self-hosted weights under a declared Apache-2.0 licence with no clawback clause, standard Llama architecture files and no custom code in the repos (ev-licence, ev-hf-platform).

How this scores
Not a scored AOI dimension. For a self-hosted model, data-control is a structural property of running the weights yourself, strong by default unless the model phones home or the licence claws back rights. For a hosted API this factor is the retention + train-on-inputs + residency read, scored from the binding terms.
What this means for adoptionYou own the use of the weights: Apache-2.0 is declared, ungated, commercial use is allowed and your data stays on your hardware. You do not own a deployable German assistant, because these are small base models with no safety evaluation, you cannot rebuild them without the unreleased training code, and every benchmark is the publisher's own. The level rises if the group ships a LICENSE file, releases its training code and configs, and settles the Llama 3.3 question on the classifiers.

Sources

The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.

Model cardread2026-11-19
The Boldt-DC-350M card declares Apache-2.0 in its metadata only, says the Boldt series is 'trained from scratch' (a publisher claim), calls the model a base model not optimised for chat, and says no systematic toxicity, bias or stereotype evaluation was done.
Model cardread2026-11-19
The Boldt-DC-1B card declares Apache-2.0 in its metadata only, gives a 2,048-token context and 200B training tokens of the German Dense-Core subset, reports a 44.05 average on the publisher's German benchmark suite, and carries the same no-safety-evaluation statement.
Licenceread2026-11-19
Both repos declare 'license: apache-2.0' in front matter and are not gated.
Technical_reportread2026-11-19
The EMNLP 2026 paper 'Repetition over Diversity' (arXiv 2604.28075, Aynetdinov, Haller and Akbik, Humboldt-Universitaet zu Berlin) describes training models 'from scratch' on filtered German FineWeb-2.
Documentationread2026-11-19
Boldt/fineweb2-de-dc is a public ODC-By 1.0 dataset of 24.4M German documents (28B tokens) subject to Common Crawl's Terms of Use, selected by three released classifiers trained on Llama-3.3-70B-Instruct labels; the classifier repos have no README, no licence and no LICENSE file.
Terms of serviceread2026-11-19
Llama 3.3 Community License section 1.b.i says that anyone who uses Llama Materials or any outputs of them to train or improve a distributed AI model shall include 'Llama' at the beginning of that model's name.
Documentationread2026-11-19
The Boldt organisation is not verified on Hugging Face and has two members.
Model cardread2026-11-19
Sibling releases differ from the DC models.
Vendor announcementread2026-11-19
The Hugging Face organisation README and the project page name the Chair of Machine Learning at Humboldt-Universitaet zu Berlin as the developer and describe Boldt as 'state-of-the-art' and 'top-tier'; these are marketing statements.
Third-party analysisread2026-11-19
No independent benchmark, leaderboard entry or safety report for any Boldt model was found; the only outside artefacts are community fine-tunes on the Hugging Face Hub.