Exegen Qwen14b v2 LoRA
QLoRA adapter for exegetical generation (terse Sanskrit → expansive English commentary), fine-tuned on the expanded ExeGen train set.
Source du modèle
- Type
- LoRA
- Taille du fichier
- 131 MB
- Modèle de base
- Qwen/Qwen2.5-14B-Instruct
- Version
- vadapter_model
- Mis à jour
- 5 oct. 2026
- Téléchargements
- 6
- Source
- Hugging Face
Description de la source
ExeGen Qwen2.5-14B v2 — QLoRA (2021-expanded)
QLoRA adapter for exegetical generation (terse Sanskrit → expansive English commentary), fine-tuned on the expanded ExeGen train set.
- Base: Qwen/Qwen2.5-14B-Instruct (4-bit nf4, QLoRA r=16, α=32, 3 epochs)
- Train: 1,594 pairs (704 original + 890 new verse→oral-commentary pairs from Mark Dyczkowski's 2021 Tantrāloka ch5/7/8/9/10 lectures), vs the v1 adapter's 704.
- Trained on Brev (7×H100 DDP, ~9 min) — successor to
exegen-qwen14b-lora.
Usage
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-14B-Instruct", device_map="auto")
model = PeftModel.from_pretrained(base, "Anamavajra-Labs/exegen-qwen14b-v2-lora")
Sources
1 source
Sources
1 sourceVérifié 23 août
Artefacts du modèle
1 artefact
Artefacts du modèle
1 artefactadapter_model.safetensors
safetensors · 131 MB · SHA-256 25167fc9e3b9…711d · Hugging Face
Download fileExtraits de sources
2 extraits
Extraits de sources
2 extraits--- license: apache-2.0 base_model: Qwen/Qwen2.5-14B-Instruct library_name: peft language: [sa, en] tags: [sanskrit, exegetical-generation, kashmir-shaivism, tantraloka, kubjika, lora, qlora, nlp, indology] datasets: [Anamavajra-Labs/tantraloka-dyczkowski-raw] --- # ExeGen Qwen2.5-14B v2 — QLoRA (2021-expanded) QLoRA adapter for **exegetical generation** (terse Sanskrit → expansive English commentary), fine-tuned on the **expanded** ExeGen train set. - **Base:** Qwen/Qwen2.5-14B-Instruct (4-bit nf4, QLoRA r=16, α=32, 3 epochs) - **Train:** 1,594 pairs (704 original + **890 new verse→oral-commentary pairs** from Mark Dyczkowski's 2021 Tantrāloka ch5/7/8/9/10 lectures), vs the v1 adapter's 704. - **Trained on Brev** (7×H100 DDP, ~9 min) — successor to [`exegen-qwen14b-lora`](https://huggingface.co/Anamavajra-Labs/exegen-qwen14b-lora). ## Usage ```python from peft import PeftModel from transformers import AutoModelForCausalLM, AutoTokenizer base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-14B-Instruct", device_map="auto") model = PeftModel.from_pretrained(base, "Anamavajra-Labs/exegen-qwen14b-v2-lora") ``` Related: 📄 [paper](https://huggingface.co/Anamavajra-Labs/exegetical-generation-paper) · 🗂️ [corpus](https://huggingface.co/datasets/Anamavajra-Labs/tantraloka-dyczkowski-raw)
Source context: 6 downloads · 0 likes · Library peft · Repo Anamavajra-Labs/exegen-qwen14b-v2-lora
