Unsloth Model Card
Browse files
README.md
CHANGED
|
@@ -1,181 +1,21 @@
|
|
| 1 |
---
|
|
|
|
| 2 |
tags:
|
|
|
|
|
|
|
| 3 |
- unsloth
|
|
|
|
|
|
|
|
|
|
|
|
|
| 4 |
---
|
| 5 |
-
# ModelCard – Gemma-3N-2B **Bilanzierungs-Experte**
|
| 6 |
|
| 7 |
-
|
| 8 |
-
**Modellname:** `unsloth/gemma-3n-E2B-it`
|
| 9 |
-
**Basis-Modell:** Google Gemma-3n-2B-it
|
| 10 |
-
**Finetuning-Framework:** Unsloth
|
| 11 |
-
**Domäne:** deutsche Bilanzierung & Rechnungswesen
|
| 12 |
|
| 13 |
-
|
|
|
|
|
|
|
| 14 |
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
## Trainingsdaten
|
| 18 |
-
| Merkmal | Wert |
|
| 19 |
-
|---------|------|
|
| 20 |
-
| Beispiele | 932 Frage-Antwort-Dialoge |
|
| 21 |
-
| Sprache | Deutsch |
|
| 22 |
-
| Fachgebiet| Bilanzierung, Jahresabschluss, Steuerrecht |
|
| 23 |
-
| Format | Strukturierte Conversational-Paare |
|
| 24 |
-
|
| 25 |
-
### Themenabdeckung
|
| 26 |
-
- Bilanzaufbau, Bilanzarten
|
| 27 |
-
- Anlage- und Umlaufvermögen, Finanzanlagen
|
| 28 |
-
- Eigenkapital, Jahresüberschuss/-fehlbetrag
|
| 29 |
-
- Handels- vs. Steuerbilanz, latente Steuern
|
| 30 |
-
- Bilanztheorien (statisch, dynamisch, organisch)
|
| 31 |
-
- Stille Rücklagen, Ausweis eigener Anteile
|
| 32 |
-
|
| 33 |
-
---
|
| 34 |
-
|
| 35 |
-
## Training-Konfiguration
|
| 36 |
-
| Parameter | Wert |
|
| 37 |
-
|-----------|------|
|
| 38 |
-
| GPU | NVIDIA T4 (16 GB VRAM) |
|
| 39 |
-
| Trainingsdauer | 85 min |
|
| 40 |
-
| Peak-VRAM | 10.99 GB |
|
| 41 |
-
|
| 42 |
-
```
|
| 43 |
-
|
| 44 |
-
per_device_train_batch_size = 1
|
| 45 |
-
gradient_accumulation_steps = 6
|
| 46 |
-
num_train_epochs = 4
|
| 47 |
-
learning_rate = 2e-5
|
| 48 |
-
warmup_steps = 50
|
| 49 |
-
total_training_steps = 624
|
| 50 |
-
optimizer = "adamw_8bit"
|
| 51 |
-
lr_scheduler_type = "cosine"
|
| 52 |
-
|
| 53 |
-
```
|
| 54 |
-
|
| 55 |
-
**Ergebnis-Metriken**
|
| 56 |
-
• Finaler Training-Loss: 1.27
|
| 57 |
-
• Effektive Batch Size: 6
|
| 58 |
-
• Trainierbare Parameter: 10.6 M (0.19 % des Modells)
|
| 59 |
-
|
| 60 |
-
---
|
| 61 |
-
|
| 62 |
-
## Nutzung
|
| 63 |
-
|
| 64 |
-
### Installation
|
| 65 |
-
```
|
| 66 |
-
|
| 67 |
-
pip install unsloth transformers torch
|
| 68 |
-
|
| 69 |
-
```
|
| 70 |
-
|
| 71 |
-
### Modell laden
|
| 72 |
-
```
|
| 73 |
-
|
| 74 |
-
from unsloth import FastLanguageModel
|
| 75 |
-
|
| 76 |
-
model, tokenizer = FastLanguageModel.from_pretrained(
|
| 77 |
-
model_name="path/to/your/model",
|
| 78 |
-
max_seq_length=2048,
|
| 79 |
-
load_in_4bit=True,
|
| 80 |
-
)
|
| 81 |
-
FastLanguageModel.for_inference(model)
|
| 82 |
-
|
| 83 |
-
```
|
| 84 |
-
|
| 85 |
-
### Prompt-Beispiel
|
| 86 |
-
```
|
| 87 |
-
|
| 88 |
-
template = """
|
| 89 |
-
Below is an instruction that describes a task, paired with an input that provides further context.
|
| 90 |
-
Write a response that appropriately completes the request.
|
| 91 |
-
|
| 92 |
-
### Instruction:
|
| 93 |
-
|
| 94 |
-
{}
|
| 95 |
-
|
| 96 |
-
### Input:
|
| 97 |
-
|
| 98 |
-
{}
|
| 99 |
-
|
| 100 |
-
### Response:
|
| 101 |
-
|
| 102 |
-
{}
|
| 103 |
-
"""
|
| 104 |
-
|
| 105 |
-
prompt = template.format(
|
| 106 |
-
"Du bist Experte für deutsches Bilanzwesen. Beantworte präzise:",
|
| 107 |
-
"Was ist eine Bilanz und welche Informationen liefert sie?",
|
| 108 |
-
"",
|
| 109 |
-
)
|
| 110 |
-
|
| 111 |
-
inputs = tokenizer([prompt], return_tensors="pt").to("cuda")
|
| 112 |
-
outputs = model.generate(
|
| 113 |
-
**inputs,
|
| 114 |
-
max_new_tokens=512,
|
| 115 |
-
temperature=1.0,
|
| 116 |
-
top_p=0.95,
|
| 117 |
-
top_k=64,
|
| 118 |
-
)
|
| 119 |
-
print(tokenizer.decode(outputs, skip_special_tokens=True))
|
| 120 |
-
|
| 121 |
-
```
|
| 122 |
-
|
| 123 |
-
**Empfohlene Inferenz-Parameter**
|
| 124 |
-
Temperature 1.0 · Top-p 0.95 · Top-k 64 · Max 512–1 024 Tokens
|
| 125 |
-
|
| 126 |
-
---
|
| 127 |
-
|
| 128 |
-
## Stärken
|
| 129 |
-
- **Fachspezifische Terminologie**: HGB- und EStG-konform.
|
| 130 |
-
- **Strukturierte Antworten**: klar gegliedert, didaktisch.
|
| 131 |
-
- **Ressourcenschonend**: 2 B-Parameter + 4-bit-Quantisierung → geringe VRAM-Anforderungen.
|
| 132 |
-
|
| 133 |
-
### Typische Einsatzgebiete
|
| 134 |
-
Aus- & Weiterbildung, erste fachliche Einschätzungen, automatisierte Q&A-Systeme im Finanz-Reporting.
|
| 135 |
-
|
| 136 |
-
---
|
| 137 |
-
|
| 138 |
-
## Limitierungen
|
| 139 |
-
- **Domänenfokus**: außerhalb des Bilanzrechts eingeschränkte Genauigkeit.
|
| 140 |
-
- **Keine Rechtsberatung**: ersetzt keine Steuer- oder Wirtschaftsprüfer.
|
| 141 |
-
- **Datenstand**: Fachinhalte bis Trainingszeitpunkt; spätere Gesetzesänderungen nicht abgedeckt.
|
| 142 |
-
|
| 143 |
-
---
|
| 144 |
-
|
| 145 |
-
## Technische Details
|
| 146 |
-
- Quantisierung: 4-Bit (BitsAndBytes)
|
| 147 |
-
- Kompatibel mit Unsloth & Transformers ≥ 4.40
|
| 148 |
-
- Inferenz ab 4 GB VRAM möglich.
|
| 149 |
-
|
| 150 |
-
---
|
| 151 |
-
|
| 152 |
-
## Lizenz – Apache 2.0
|
| 153 |
-
Dieses Modell wird unter der Apache License, Version 2.0 bereitgestellt.
|
| 154 |
-
Die Nutzung ist frei, auch kommerziell, bei Beibehaltung der Lizenz- und Haftungsbedingungen.
|
| 155 |
-
|
| 156 |
-
> Licensed under the Apache License, Version 2.0 (the “License”);
|
| 157 |
-
> you may not use this file except in compliance with the License.
|
| 158 |
-
> Unless required by applicable law or agreed to in writing, software
|
| 159 |
-
> distributed under the License is distributed on an “AS IS” BASIS,
|
| 160 |
-
> WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 161 |
-
> See the License for the specific language governing permissions and
|
| 162 |
-
> limitations under the License.
|
| 163 |
-
|
| 164 |
-
*Hinweis: Das zugrunde liegende Gemma-Modell unterliegt zusätzlich den Google Gemma Terms of Use.*
|
| 165 |
-
|
| 166 |
-
---
|
| 167 |
-
|
| 168 |
-
## Kontakt & Support
|
| 169 |
-
Fragen oder Feedback zum Einsatz im deutschen Bilanzwesen?
|
| 170 |
-
`<Ihre Kontaktadresse>`
|
| 171 |
-
|
| 172 |
-
---
|
| 173 |
-
|
| 174 |
-
## Haftungsausschluss
|
| 175 |
-
Dieses Modell dient ausschließlich Bildungs- und Informationszwecken und ersetzt keine professionelle Beratungsleistung. Entscheidungen basierend auf Modell-Ausgaben erfolgen auf eigenes Risiko.
|
| 176 |
-
```
|
| 177 |
-
|
| 178 |
-
<div style="text-align: center">⁂</div>
|
| 179 |
-
|
| 180 |
-
[^1]: Bitte-die-gesamte-Modelcard-als-Markdown-ausgeben.md
|
| 181 |
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
base_model: unsloth/gemma-3n-e2b-it-unsloth-bnb-4bit
|
| 3 |
tags:
|
| 4 |
+
- text-generation-inference
|
| 5 |
+
- transformers
|
| 6 |
- unsloth
|
| 7 |
+
- gemma3n
|
| 8 |
+
license: apache-2.0
|
| 9 |
+
language:
|
| 10 |
+
- en
|
| 11 |
---
|
|
|
|
| 12 |
|
| 13 |
+
# Uploaded finetuned model
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
|
| 15 |
+
- **Developed by:** ProfEngel
|
| 16 |
+
- **License:** apache-2.0
|
| 17 |
+
- **Finetuned from model :** unsloth/gemma-3n-e2b-it-unsloth-bnb-4bit
|
| 18 |
|
| 19 |
+
This gemma3n model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 20 |
|
| 21 |
+
[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)
|