Upload lora_config.yaml with huggingface_hub
Browse files- lora_config.yaml +42 -0
lora_config.yaml
ADDED
|
@@ -0,0 +1,42 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# MLX-LM LoRA config for fine-tuning VibeThinker-3B into a bug-bounty triage model.
|
| 2 |
+
model: "WeiboAI/VibeThinker-3B"
|
| 3 |
+
train: true
|
| 4 |
+
|
| 5 |
+
# Data dir must contain train.jsonl and valid.jsonl (chat / messages format).
|
| 6 |
+
data: "data/sft"
|
| 7 |
+
|
| 8 |
+
fine_tune_type: lora
|
| 9 |
+
optimizer: adamw
|
| 10 |
+
seed: 13
|
| 11 |
+
|
| 12 |
+
# Apply LoRA to all 36 transformer blocks for richer adaptation (M5 Max has headroom).
|
| 13 |
+
num_layers: 36
|
| 14 |
+
|
| 15 |
+
batch_size: 8
|
| 16 |
+
iters: 2000
|
| 17 |
+
max_seq_length: 2048
|
| 18 |
+
grad_checkpoint: true
|
| 19 |
+
|
| 20 |
+
learning_rate: 1.0e-4
|
| 21 |
+
# Only learn from the assistant turn, not the (long) system+report prompt.
|
| 22 |
+
mask_prompt: true
|
| 23 |
+
|
| 24 |
+
steps_per_report: 10
|
| 25 |
+
steps_per_eval: 200
|
| 26 |
+
val_batches: 25
|
| 27 |
+
save_every: 200
|
| 28 |
+
|
| 29 |
+
adapter_path: "adapters"
|
| 30 |
+
|
| 31 |
+
lora_parameters:
|
| 32 |
+
keys:
|
| 33 |
+
- "self_attn.q_proj"
|
| 34 |
+
- "self_attn.k_proj"
|
| 35 |
+
- "self_attn.v_proj"
|
| 36 |
+
- "self_attn.o_proj"
|
| 37 |
+
- "mlp.gate_proj"
|
| 38 |
+
- "mlp.up_proj"
|
| 39 |
+
- "mlp.down_proj"
|
| 40 |
+
rank: 16
|
| 41 |
+
scale: 20.0
|
| 42 |
+
dropout: 0.0
|