Views
No views yet
<|im_start|> and <|im_end|>)<|im_start|>system
{system_message}<|im_end|>
<|im_start|>user
{user_message}<|im_end|>
<|im_start|>assistant1from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer
2import torch
3
4model_path = "Felladrin/TinyMistral-248M-Chat-v3"
5device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
6tokenizer = AutoTokenizer.from_pretrained(model_path)
7model = AutoModelForCausalLM.from_pretrained(model_path).to(device)
8streamer = TextStreamer(tokenizer)
9messages = [
10 {
11 "role": "system",
12 "content": "You are a highly knowledgeable and friendly assistant. Your goal is to understand and respond to user inquiries with clarity. Your interactions are always respectful, helpful, and focused on delivering the most accurate information to the user.",
13 },
14 {
15 "role": "user",
16 "content": "Hey! Got a question for you!",
17 },
18 {
19 "role": "assistant",
20 "content": "Sure! What's it?",
21 },
22 {
23 "role": "user",
24 "content": "What are some potential applications for quantum computing?",
25 },
26]
27prompt = tokenizer.apply_chat_template(
28 messages, tokenize=False, add_generation_prompt=True
29)
30inputs = tokenizer(prompt, return_tensors="pt").to(device)
31model.generate(
32 inputs.input_ids,
33 attention_mask=inputs.attention_mask,
34 max_length=tokenizer.model_max_length,
35 streamer=streamer,
36 eos_token_id=tokenizer.eos_token_id,
37 pad_token_id=tokenizer.pad_token_id,
38 do_sample=True,
39 temperature=0.6,
40 top_p=0.8,
41 top_k=0,
42 min_p=0.1,
43 typical_p=0.2,
44 repetition_penalty=1.176,
45)| Hyperparameter | Value |
|---|---|
| Learning rate | 2e-5 |
| Total train batch size | 32 |
| Max. sequence length | 2048 |
| Weight decay | 0.01 |
| Warmup ratio | 0.1 |
| NEFTune Noise Alpha | 5 |
| Optimizer | Adam with betas=(0.9,0.999) and epsilon=1e-08 |
| Scheduler | cosine |
| Seed | 42 |
| Parameter | Value |
|---|---|
| Dataset | HuggingFaceH4/ultrafeedback_binarized |
| Learning rate | 1e-06 |
| Train batch size | 4 |
| Eval batch size | 8 |
| Seed | 42 |
| Distributed type | multi-GPU |
| Number of devices | 8 |
| Gradient accumulation steps | 4 |
| Total train batch size | 128 |
| Total eval batch size | 64 |
| Optimizer | adamw_8bit with betas=(0.9,0.999) and epsilon=1e-08 |
| LR scheduler type | cosine |
| LR scheduler warmup ratio | 0.1 |
| Number of epochs | 2.0 |
1llamafactory-cli train \
2 --stage dpo \
3 --do_train True \
4 --model_name_or_path ~/TinyMistral-248M-Chat \
5 --preprocessing_num_workers $(python -c "import os; print(max(1, os.cpu_count() - 2))") \
6 --dataloader_num_workers $(python -c "import os; print(max(1, os.cpu_count() - 2))") \
7 --finetuning_type full \
8 --template default \
9 --flash_attn auto \
10 --enable_liger_kernel True \
11 --dataset_dir data \
12 --dataset ultrafeedback \
13 --cutoff_len 1024 \
14 --learning_rate 1e-6 \
15 --num_train_epochs 2.0 \
16 --per_device_train_batch_size 4 \
17 --gradient_accumulation_steps 4 \
18 --lr_scheduler_type cosine \
19 --max_grad_norm 1.0 \
20 --logging_steps 10 \
21 --save_steps 50 \
22 --save_total_limit 1 \
23 --warmup_ratio 0.1 \
24 --packing False \
25 --report_to none \
26 --output_dir ~/TinyMistral-248M-Chat-v3 \
27 --pure_bf16 True \
28 --plot_loss True \
29 --trust_remote_code True \
30 --ddp_timeout 180000000 \
31 --include_tokens_per_second True \
32 --include_num_input_tokens_seen True \
33 --optim adamw_8bit \
34 --pref_beta 0.5 \
35 --pref_ftx 0 \
36 --pref_loss simpo \
37 --gradient_checkpointing True