Twscholar Lm Sft 1.5b LoRA R64
Part of a three-line scale comparison (0.5B full fine-tune / 1.5B LoRA / 7B QLoRA) documenting VRAM and quality trade-offs on a single 12GB consumer GPU. This is the 1.5B point — not the recommended model for actual...
Modellquelle
- Art
- LoRA
- Dateigröße
- 282 MB
- Basismodell
- Qwen/Qwen2.5-1.5B-Instruct
- Version
- vFP32
- Aktualisiert
- 3. Nov. 2026
- Downloads
- 0
- Quelle
- Hugging Face
Quellenbeschreibung
twscholar-lm — 1.5B LoRA (scale-comparison checkpoint)
Part of a three-line scale comparison (0.5B full fine-tune / 1.5B LoRA /
7B QLoRA) documenting VRAM and quality trade-offs on a single 12GB
consumer GPU. This is the 1.5B point — not the recommended model for
actual use (see twscholar-lm-dpo-7b-final), published for the
comparison's reproducibility.
- LoRA r=64, alpha=128, 3 epochs
- Peak VRAM during training: 6.07GB
- Full comparison table:
results/m2_sft_comparison.jsonin the project repo — https://github.com/oscarlin778/twscholar-lm - License: cc-by-nc-4.0 (inherited from the training dataset)
Quellen
1 Quelle
Quellen
1 QuelleVerifiziert 3. Nov.
Download files
Repository files
File inventory only; completeness is unverified.
- adapter_config.json
- adapter_model.safetensors
- checkpoint-30/adapter_config.json
- checkpoint-30/adapter_model.safetensors
- checkpoint-30/tokenizer.json
- checkpoint-30/tokenizer_config.json
- checkpoint-30/trainer_state.json
- checkpoint-60/adapter_config.json
- checkpoint-60/adapter_model.safetensors
- checkpoint-60/tokenizer.json
- checkpoint-60/tokenizer_config.json
- checkpoint-60/trainer_state.json
- checkpoint-90/adapter_config.json
- checkpoint-90/adapter_model.safetensors
- checkpoint-90/tokenizer.json
- checkpoint-90/tokenizer_config.json
- checkpoint-90/trainer_state.json
- run_metrics.json
- tokenizer.json
- tokenizer_config.json
Modellartefakte
2 Artefakte
Modellartefakte
2 Artefakteadapter_model.safetensors
safetensors · FP32 (tensor header) · 282 MB · SHA-256 650dd34ae54e…f5e6 · Hugging Face
Download filecheckpoint-90/adapter_model.safetensors
safetensors · FP32 (tensor header) · 282 MB · SHA-256 650dd34ae54e…f5e6 · Hugging Face
Download fileQuellenauszüge
2 Auszüge
Quellenauszüge
2 Auszüge--- license: cc-by-nc-4.0 base_model: Qwen/Qwen2.5-1.5B-Instruct tags: - lora - sft - traditional-chinese - academic-writing language: - zh --- # twscholar-lm — 1.5B LoRA (scale-comparison checkpoint) Part of a three-line scale comparison (0.5B full fine-tune / 1.5B LoRA / 7B QLoRA) documenting VRAM and quality trade-offs on a single 12GB consumer GPU. This is the 1.5B point — not the recommended model for actual use (see `twscholar-lm-dpo-7b-final`), published for the comparison's reproducibility. - LoRA r=64, alpha=128, 3 epochs - Peak VRAM during training: 6.07GB - Full comparison table: `results/m2_sft_comparison.json` in the project repo — https://github.com/oscarlin778/twscholar-lm - License: cc-by-nc-4.0 (inherited from the training dataset)
Source context: 0 downloads · 0 likes · Repo Yoda6131027/twscholar-lm-sft-1.5b-lora-r64
