Update README.md
Browse files
README.md
CHANGED
@@ -14,7 +14,8 @@ pipeline_tag: text-generation
|
|
14 |
|
15 |
# This is highly experimental and should be viewed as purely testing right now. Jamba has been very hard to train but I wanted to see how it did on one of the best datasets we have access to. I believe in transparent development so all *best* working iterations, even if they are a bit wonky, will be pushed here.
|
16 |
|
17 |
-
|
|
|
18 |
|
19 |
*There's been limited testing so no example outputs yet*
|
20 |
|
@@ -41,7 +42,7 @@ lora_config = LoraConfig(
|
|
41 |
r=8,
|
42 |
lora_alpha=16,
|
43 |
target_modules=["embed_tokens", "x_proj", "in_proj", "out_proj"],
|
44 |
-
lora_dropout=0.
|
45 |
task_type="CAUSAL_LM",
|
46 |
bias="none"
|
47 |
)
|
@@ -54,19 +55,19 @@ trainer = SFTTrainer(
|
|
54 |
tokenizer=tokenizer,
|
55 |
args=TrainingArguments(
|
56 |
num_train_epochs=1,
|
57 |
-
lr_scheduler_type='
|
58 |
-
learning_rate=
|
59 |
per_device_train_batch_size=1,
|
60 |
gradient_accumulation_steps=8,
|
61 |
gradient_checkpointing=True,
|
62 |
warmup_steps=10,
|
63 |
-
weight_decay=0.
|
64 |
fp16=not torch.cuda.is_bf16_supported(),
|
65 |
bf16=torch.cuda.is_bf16_supported(),
|
66 |
logging_steps=1,
|
67 |
-
save_steps=
|
68 |
output_dir="outputs",
|
69 |
-
optim="
|
70 |
seed=42,
|
71 |
),
|
72 |
)
|
|
|
14 |
|
15 |
# This is highly experimental and should be viewed as purely testing right now. Jamba has been very hard to train but I wanted to see how it did on one of the best datasets we have access to. I believe in transparent development so all *best* working iterations, even if they are a bit wonky, will be pushed here.
|
16 |
|
17 |
+
---
|
18 |
+
# New training underway! Thanks to the generous insights provided by **lightblue/Jamba-v0.1-chat-multilingual**, the new training is going much better. We should hopefully have a decently trained Jamaba-Open-Hermes model for general use and experimentation.
|
19 |
|
20 |
*There's been limited testing so no example outputs yet*
|
21 |
|
|
|
42 |
r=8,
|
43 |
lora_alpha=16,
|
44 |
target_modules=["embed_tokens", "x_proj", "in_proj", "out_proj"],
|
45 |
+
lora_dropout=0.05,
|
46 |
task_type="CAUSAL_LM",
|
47 |
bias="none"
|
48 |
)
|
|
|
55 |
tokenizer=tokenizer,
|
56 |
args=TrainingArguments(
|
57 |
num_train_epochs=1,
|
58 |
+
lr_scheduler_type='cosine',
|
59 |
+
learning_rate=0.0002,
|
60 |
per_device_train_batch_size=1,
|
61 |
gradient_accumulation_steps=8,
|
62 |
gradient_checkpointing=True,
|
63 |
warmup_steps=10,
|
64 |
+
weight_decay=0.01,
|
65 |
fp16=not torch.cuda.is_bf16_supported(),
|
66 |
bf16=torch.cuda.is_bf16_supported(),
|
67 |
logging_steps=1,
|
68 |
+
save_steps=200,
|
69 |
output_dir="outputs",
|
70 |
+
optim="adamw_8bit",
|
71 |
seed=42,
|
72 |
),
|
73 |
)
|