IlyaGusev commited on
Commit
fb506de
1 Parent(s): b9cbc5c

New version

Browse files
adapter_config.json CHANGED
@@ -1,5 +1,5 @@
1
  {
2
- "base_model_name_or_path": "models/llama-30b",
3
  "bias": "none",
4
  "fan_in_fan_out": false,
5
  "inference_mode": true,
@@ -16,4 +16,4 @@
16
  "o_proj"
17
  ],
18
  "task_type": "CAUSAL_LM"
19
- }
 
1
  {
2
+ "base_model_name_or_path": "huggyllama/llama-30b",
3
  "bias": "none",
4
  "fan_in_fan_out": false,
5
  "inference_mode": true,
 
16
  "o_proj"
17
  ],
18
  "task_type": "CAUSAL_LM"
19
+ }
adapter_model.bin CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:60a82a49682551b7836176ec9f08a7d8f4ec97c090ae11f4f14383658bbac385
3
  size 204646285
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9efdee038d78af46f77465a09f2c97fb3d7d3fb86582140567fa7a8a1b38a858
3
  size 204646285
generation_config.json CHANGED
@@ -2,12 +2,12 @@
2
  "pad_token_id": 0,
3
  "bos_token_id": 1,
4
  "eos_token_id": 2,
5
- "temperature": 1.0,
6
  "top_p": 0.9,
7
  "top_k": 30,
8
  "do_sample": true,
9
  "max_new_tokens": 1536,
10
- "num_beams": 2,
11
  "repetition_penalty": 1.1,
12
- "no_repeat_ngram_size": 8
13
  }
 
2
  "pad_token_id": 0,
3
  "bos_token_id": 1,
4
  "eos_token_id": 2,
5
+ "temperature": 0.2,
6
  "top_p": 0.9,
7
  "top_k": 30,
8
  "do_sample": true,
9
  "max_new_tokens": 1536,
10
+ "num_beams": 1,
11
  "repetition_penalty": 1.1,
12
+ "no_repeat_ngram_size": 15
13
  }
template.json ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ {
2
+ "system_prompt": "Ты — Сайга, русскоязычный автоматический ассистент. Ты разговариваешь с людьми и помогаешь им.",
3
+ "message_template": "<s>{role}\n{content}</s>\n",
4
+ "start_token_id": 1,
5
+ "end_token_id": 2,
6
+ "bot_token_id": 9225
7
+ }
training_config.json ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "trainer": {
3
+ "evaluation_strategy": "steps",
4
+ "per_device_train_batch_size": 6,
5
+ "per_device_eval_batch_size": 6,
6
+ "gradient_accumulation_steps": 21,
7
+ "eval_steps": 50,
8
+ "save_steps": 50,
9
+ "logging_steps": 5,
10
+ "learning_rate": 0.0003,
11
+ "num_train_epochs": 5,
12
+ "lr_scheduler_type": "cosine",
13
+ "warmup_steps": 30,
14
+ "fp16": true,
15
+ "bf16": false,
16
+ "torch_compile": false,
17
+ "optim": "adamw_torch"
18
+ },
19
+ "lora": {
20
+ "r": 16,
21
+ "lora_alpha": 16,
22
+ "lora_dropout": 0.05,
23
+ "bias": "none",
24
+ "target_modules": ["q_proj", "v_proj", "k_proj", "o_proj"],
25
+ "task_type": "CAUSAL_LM"
26
+ },
27
+ "load_in_8bit": true,
28
+ "only_target_loss": true,
29
+ "mode": "chat",
30
+ "templates_path": "internal_prompts/saiga_v2.json",
31
+ "model_name": "models/llama-30b",
32
+ "model_type": "causal",
33
+ "max_tokens_count": 2000
34
+ }
35
+