Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Refactoring reinforce #23

Open
wants to merge 14 commits into
base: main
Choose a base branch
from
Open

Refactoring reinforce #23

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
14 commits
Select commit Hold shift + click to select a range
1d8beb0
reinforce
ummagumm-a Aug 28, 2024
f13d100
initial
syrn1k Sep 16, 2024
f9aa5db
Merge branch 'refs/heads/main' into reinforce_refactoring
syrn1k Sep 22, 2024
afda1da
in progress
syrn1k Sep 22, 2024
3bf784c
Merge branch 'main' into reinforce_refactoring
syrn1k Sep 22, 2024
c6f4335
in progress
syrn1k Sep 22, 2024
8e19266
in progress
syrn1k Sep 27, 2024
243c1f0
in progress
syrn1k Oct 2, 2024
55054ed
in progress
syrn1k Oct 6, 2024
ae8be1b
in progress
syrn1k Oct 12, 2024
120c2a9
in progress
syrn1k Oct 14, 2024
a7ed269
in progress
syrn1k Oct 19, 2024
f5a7d1d
in progress
syrn1k Oct 19, 2024
9914764
in progress
syrn1k Oct 29, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion configs/exp/train/dpo/dpo.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -100,7 +100,7 @@
"bos_token": "<|begin_of_text|>",
"eos_token": "<|end_of_text|>"
},
"trainer_settings": {
"trainer_settings": {
"evaluation_strategy": "steps",
"per_device_train_batch_size": 1,
"per_device_eval_batch_size": 1,
Expand Down
214 changes: 214 additions & 0 deletions configs/exp/train/reinforce/reinforce.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,214 @@
{
"train_dataset_settings": {
"sources": [
{
"name": "train_hh",
"records_path": "/from_s3/datasets/train_chat_filtered.jsonl",
"sample_rate": 1.0
}
],
"prompt_template": {
"role_tag_mapping": {
"bot": "assistant",
"user": "user",
"system": "system"
},
"prefix_template": "<|start_header_id|>{role}<|end_header_id|>\n\n",
"suffix_template": "<|eot_id|>"
},
"dataset_type": "chat",
"max_tokens_count": 1024,
"only_answer_loss": true,
"keep_end": true,
"random_cut": false,
"check_max_len_enforced": true
},
"val_dataset_settings": {
"sources": [
{
"name": "val_hh",
"records_path": "/from_s3/datasets/val_chat_filtered.jsonl",
"sample_rate": 0.2
}
],
"prompt_template": {
"role_tag_mapping": {
"bot": "assistant",
"user": "user",
"system": "system"
},
"prefix_template": "<|start_header_id|>{role}<|end_header_id|>\n\n",
"suffix_template": "<|eot_id|>"
},
"dataset_type": "chat",
"max_tokens_count": 1024,
"only_answer_loss": true,
"keep_end": true,
"random_cut": false,
"check_max_len_enforced": true
},

"cherry_pick_settings": {
"generator_transformers_settings": {
"num_beams": 1,
"do_sample": true,
"num_return_sequences": 3,
"max_new_tokens": 1024,
"temperature": 0.7,
"top_p": 1.0,
"top_k": 0,
"repetition_penalty": 1.0
},
"custom_generation_settings": {
"generation_eos_token": "<|eot_id|>",
"remove_prompt": true,
"skip_special_tokens": true
},
"dataset_settings": {
"sources": [
{
"name": "cp_hh",
"records_path": "/from_s3/datasets/val_chat_filtered.jsonl",
"num_samples": 10
}
],
"prompt_template": {
"role_tag_mapping": {
"bot": "assistant",
"user": "user",
"system": "system"
},
"prefix_template": "<|start_header_id|>{role}<|end_header_id|>\n\n",
"suffix_template": "<|eot_id|>"
},
"dataset_type": "chat",
"max_tokens_count": 1024,
"random_cut": false,
"keep_end": true,
"only_answer_loss": true,
"check_max_len_enforced": true
},
"metric_settings": [
{
"type": "length",
"parameters": {
"need_average": [true, false]
}
},
{
"type": "self_bleu",
"parameters": {
"need_average": [true, false]
}
},
{
"type": "dist_n",
"parameters": {
"need_average": [true, false]
}
},
{
"type": "diversity",
"parameters": {
"need_average": [true, false]
}
},
{
"type": "perplexity",
"parameters": {
"need_average": [true, false]
}
}
]
},
"reward_model_settings": {
"model_path": "/from_s3/models/rm",
"model_type": "llama_reward_value",
"model_kwargs": {
"num_labels": 1,
"attn_implementation": "flash_attention_2"
},
"transformers_settings": {},
"check_checkpoint_is_compatible": true
},
"model_settings": {
"model_path": "/from_s3/models/policy/",
"model_type": "causal",
"transformers_settings": {},
"model_kwargs": {
"attn_implementation": "flash_attention_2"
},
"check_checkpoint_is_compatible": true
},
"tokenizer_settings": {
"use_fast": true,
"padding_side": "left"
},
"peft_settings": {
"inference_mode": false,
"r": 64,
"lora_alpha": 64,
"lora_dropout": 0.0,
"use_rslora": false,
"init_lora_weights": true,
"task_type": "CAUSAL_LM"
},
"trainer_settings": {
"evaluation_strategy": "steps",
"per_device_train_batch_size": 8,
"per_device_eval_batch_size": 8,
"gradient_accumulation_steps": 10,
"adam_beta1": 0.9,
"adam_beta2": 0.95,
"adam_epsilon": 1e-5,
"eval_steps": 50,
"save_steps": 100,
"save_strategy": "steps",
"load_best_model_at_end": false,
"logging_steps": 1,
"learning_rate": 1e-5,
"max_steps": 1001,
"lr_scheduler_type": "constant_with_warmup",
"warmup_ratio": 0.03,
"fp16": false,
"bf16": true,
"optim": "adamw_torch",
"weight_decay": 0.0,
"max_grad_norm": 2,
"save_total_limit": 11,
"dataloader_num_workers": 12,
"deepspeed": "configs/exp/deepspeed/ds_config_stage_2_reinforce.json",

"max_response_length": 1024,
"stop_token": "<|eot_id|>",
"temperature": 0.7,
"non_eos_penalty": true,
"penalty_reward_value": -1,
"clip_rewards_min": -1e8,
"clip_rewards_max": 1e8,
"whiten_rewards": false,
"kl_coef": 0.05,
"risk_coef": 0.00,
"mean_baseline_coef": 0.95,

"init_kl_coef": null,
"target_kl": null,
"adaptive_kl_k": null,
"adaptive_kl_clip_value": null,
"min_kl_coef": null,
"max_kl_coef": null,

"num_samples_for_reward_stats": 1000

},
"wandb_settings": {
"project_name": "...",
"group": "hh-llama3-8b",
"job_type": "reinforce",
"run_name": "rvm-kl_0.05",
"entity": "...",
"tags": ["reinforce", "llama3-8b", "hh", "rvm"]
},
"log_path": "train_output/hh/llama3-8b/reinforce/rvm-kl_0.05-0.01/"
}

175 changes: 175 additions & 0 deletions tests/fixtures/configs/train/reinforce/base.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,175 @@
{
"train_dataset_settings": {
"sources": [
{
"name": "train_chat",
"records_path": "tests/fixtures/datasets/chat/train_chat.jsonl",
"sample_rate": 1.0
}
],
"prompt_template": {
"role_tag_mapping": {
"bot": "assistant",
"user": "user",
"system": "system"
},
"prefix_template": "<|start_header_id|>{role}<|end_header_id|>\n\n",
"suffix_template": "<|eot_id|>"
},
"dataset_type": "chat",
"max_tokens_count": 2000,
"only_answer_loss": true
},
"val_dataset_settings": {
"sources": [
{
"name": "val_chat",
"records_path": "tests/fixtures/datasets/chat/train_chat.jsonl",
"sample_rate": 1.0
}
],
"prompt_template": {
"role_tag_mapping": {
"bot": "assistant",
"user": "user",
"system": "system"
},
"prefix_template": "<|start_header_id|>{role}<|end_header_id|>\n\n",
"suffix_template": "<|eot_id|>"
},
"dataset_type": "chat",
"max_tokens_count": 2000,
"only_answer_loss": true
},
"cherry_pick_settings": {
"generator_transformers_settings": {
"num_beams": 1,
"do_sample": false,
"stop_strings": "</RS>",
"max_new_tokens": 8
},
"custom_generation_settings": {
"skip_special_tokens": false
},
"dataset_settings": {
"sources": [
{
"name": "chat_test",
"records_path": "tests/fixtures/datasets/chat/train_chat.jsonl",
"num_samples": 2
}
],
"prompt_template": {
"role_tag_mapping": {
"bot": "assistant",
"user": "user",
"system": "system"
},
"prefix_template": "<|start_header_id|>{role}<|end_header_id|>\n\n",
"suffix_template": "<|eot_id|>"
},
"dataset_type": "chat",
"max_tokens_count": 150,
"only_answer_loss": true
},
"metric_settings": [
{
"type": "length",
"parameters": {
"need_average": [
true
]
}
},
{
"type": "kl",
"parameters": {
"need_average": [
true
],
"ref_logits_type": "sft"
}
},
{
"type": "kl",
"parameters": {
"need_average": [
true
],
"ref_logits_type": "reference"
}
}
]
},
"reward_model_settings": {
"model_path": "tests/fixtures/models/llama2_tiny",
"model_type": "seq_cls",
"model_kwargs": {
"num_labels": 1
},
"transformers_settings": {}
},
"model_settings": {
"model_path": "tests/fixtures/models/llama2_tiny",
"model_type": "causal",
"transformers_settings": {},
"embeddings_initialization_strategy": {
"<|start_header_id|>": "<s>",
"<|end_header_id|>": "</s>",
"<|eot_id|>": "</s>"
}
},
"tokenizer_settings": {
"tokenizer_kwargs": {
"padding_side": "left"
}
},
"trainer_settings": {
"evaluation_strategy": "steps",
"per_device_train_batch_size": 8,
"per_device_eval_batch_size": 8,
"gradient_accumulation_steps": 10,
"adam_beta1": 0.9,
"adam_beta2": 0.95,
"adam_epsilon": 0.00001,
"eval_steps": 50,
"save_steps": 100,
"save_strategy": "steps",
"load_best_model_at_end": false,
"logging_steps": 1,
"learning_rate": 0.00001,
"max_steps": 1001,
"lr_scheduler_type": "constant_with_warmup",
"warmup_ratio": 0.03,
"fp16": false,
"bf16": false,
"optim": "adamw_torch",
"weight_decay": 0.0,
"max_grad_norm": 2,
"save_total_limit": 11,
"dataloader_num_workers": 12,
"stop_token": "<|eot_id|>",
"temperature": 0.7,
"non_eos_penalty": true,
"penalty_reward_value": -1,
"clip_rewards_min": -1e+8,
"clip_rewards_max": 1e+8,
"whiten_rewards": false,
"kl_coef": 0.05,
"mean_baseline_coef": 0.95,
"num_samples_for_reward_stats": 1000,
"no_cuda": true
},
"special_tokens_settings": {
"bos_token": "<s>",
"eos_token": "</s>",
"pad_token": "<unk>"
},
"logging_settings": {
"project_name": "alignment",
"run_name": "sft",
"entity": "turbo-alignment"
},
"seed": 0,
"log_path": "train_output"
}
Loading
Loading