---

license: other

license_name: lfm1.0

license_link: LICENSE

base_model: LiquidAI/LFM2.5-1.2B-Instruct

language:

• de
• en
• fr
• ar

pipeline_tag: text-generation

tags:

• gguf
• lfm2
• german
• language-learning
• on-device

---

Laura — LFM2.5 1.2B, fine-tuned (GGUF)

Laura is the conversation partner in an offline German course app for adult

beginners (A1–A2). This model is what the app runs on the phone, CPU only.

This is a modified version of LiquidAI/LFM2.5-1.2B-Instruct.

Changes made: a LoRA adapter (rank 16, all linear layers, 2 epochs) was

trained on ~3,000 short A1–A2 German practice conversations, merged into the

weights, and the result converted to GGUF and quantised to Q4_K_M with

llama.cpp. Distributed under the base model's LFM Open License v1.0 — see

LICENSE.

What it was trained to do

• Answer in simple German, one or two short sentences, then ask one new question.
• Stay in one fixed character (Laura: from France, lives in Berlin, a cook).
• Correct mistakes only by saying the right form back, never by lecturing.
• When the learner asks what a word means, give one short sentence in the

learner's language (English, French or Arabic), using the dictionary meaning

the app supplies.

How it expects to be prompted

A one-line system prompt, then the conversation:

Du bist Laura. Niveau A1.1. Lernersprache: Französisch. Kapitel: Guten Tag!. Wörter: der Name, das Land, …

For a "Was bedeutet …?" question, the app appends the dictionary meaning to

the learner's message:

Was bedeutet Termin?

(Wörterbuch: der Termin = الموعد)

Sampling: temperature 0.1, top_k 50, top_p 1.0.

Results

Graded by hand on 70 held-out conversations (348 replies) the model never saw:

| | base LFM2.5 1.2B | this model |

|---|---|---|

| Replies that respond sensibly, in character | 52% | 89% |

| Grammatical, natural German | 88% | 97% |

Files

| file | size | sha256 |

|---|---|---|

| laura-lfm2.5-1.2b-Q4_K_M.gguf | 730,898,432 bytes | f25d0858ddced1ebf1fdef109560958c3d50eef150965212d4f13c5921cda3f1 |