New version

adapter_config.json

@@ -1,5 +1,5 @@
 {
-  "base_model_name_or_path": "models/llama-30b",
+  "base_model_name_or_path": "huggyllama/llama-30b",
   "bias": "none",
   "fan_in_fan_out": false,
   "inference_mode": true,
@@ -16,4 +16,4 @@
     "o_proj"
   ],
   "task_type": "CAUSAL_LM"
-}
+}

adapter_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:60a82a49682551b7836176ec9f08a7d8f4ec97c090ae11f4f14383658bbac385
+oid sha256:9efdee038d78af46f77465a09f2c97fb3d7d3fb86582140567fa7a8a1b38a858
 size 204646285

generation_config.json

@@ -2,12 +2,12 @@
     "pad_token_id": 0,
     "bos_token_id": 1,
     "eos_token_id": 2,
-    "temperature": 1.0,
+    "temperature": 0.2,
     "top_p": 0.9,
     "top_k": 30,
     "do_sample": true,
     "max_new_tokens": 1536,
-    "num_beams": 2,
+    "num_beams": 1,
     "repetition_penalty": 1.1,
-    "no_repeat_ngram_size": 8
+    "no_repeat_ngram_size": 15
 }

template.json

@@ -0,0 +1,7 @@
+{
+    "system_prompt": "Ты — Сайга, русскоязычный автоматический ассистент. Ты разговариваешь с людьми и помогаешь им.",
+    "message_template": "<s>{role}\n{content}</s>\n",
+    "start_token_id": 1,
+    "end_token_id": 2,
+    "bot_token_id": 9225
+}

training_config.json

@@ -0,0 +1,35 @@
+{
+    "trainer": {
+        "evaluation_strategy": "steps",
+        "per_device_train_batch_size": 6,
+        "per_device_eval_batch_size": 6,
+        "gradient_accumulation_steps": 21,
+        "eval_steps": 50,
+        "save_steps": 50,
+        "logging_steps": 5,
+        "learning_rate": 0.0003,
+        "num_train_epochs": 5,
+        "lr_scheduler_type": "cosine",
+        "warmup_steps": 30,
+        "fp16": true,
+        "bf16": false,
+        "torch_compile": false,
+        "optim": "adamw_torch"
+    },
+    "lora": {
+        "r": 16,
+        "lora_alpha": 16,
+        "lora_dropout": 0.05,
+        "bias": "none",
+        "target_modules": ["q_proj", "v_proj", "k_proj", "o_proj"],
+        "task_type": "CAUSAL_LM"
+    },
+    "load_in_8bit": true,
+    "only_target_loss": true,
+    "mode": "chat",
+    "templates_path": "internal_prompts/saiga_v2.json",
+    "model_name": "models/llama-30b",
+    "model_type": "causal",
+    "max_tokens_count": 2000
+}
+