Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: peft
|
| 3 |
+
base_model: Qwen/Qwen3-14B
|
| 4 |
+
tags: [loracle, lora, lots-of-loras, weight-reading]
|
| 5 |
+
---
|
| 6 |
+
|
| 7 |
+
# LoL LoRAcle v2 (oneq)
|
| 8 |
+
|
| 9 |
+
Fine-tuned from [ceselder/loracle-pretrain-v7-sweep-A-oneq-final-step3120](https://huggingface.co/ceselder/loracle-pretrain-v7-sweep-A-oneq-final-step3120) on Sonnet-4.6-generated Q/A about the 1,170 [Lots-of-LoRAs Qwen3-14B](https://huggingface.co/ceselder/lots-of-loras-qwen3-14b) tasks. 1 random Q/A per task, mirroring the v7 oneq regime.
|
| 10 |
+
|
| 11 |
+
## Training
|
| 12 |
+
- Q/A: [ceselder/lol-loracle-qa-v1](https://huggingface.co/datasets/ceselder/lol-loracle-qa-v1) (8 question types per task, 1 random sampled with seed=42 per task)
|
| 13 |
+
- 1129 train tasks (40 holdout), 141 steps, lr=1e-5 linear, grad_accum=8
|
| 14 |
+
- val_loss: 2.4669 (step 0, v7 baseline) -> 1.5108 (step 140 final)
|
| 15 |
+
- Cross-LoRA: matched=1.5107, crossed=1.7976 (gap=0.2869) -- loracle IS conditioning on direction tokens
|
| 16 |
+
- wandb: https://wandb.ai/adamkarvonen/lora-oracles/runs/pk6jmmne
|