Initial commit (clean history)HEADmain

9 files changed, 208 insertions, 0 deletions

diff --git a/configs/base.yaml b/configs/base.yaml
new file mode 100644
index 0000000..e69de29
--- /dev/null
+++ b/configs/base.yaml
diff --git a/configs/local_models.yaml b/configs/local_models.yaml
new file mode 100644
index 0000000..13c7fcc
--- /dev/null
+++ b/configs/local_models.yaml
@@ -0,0 +1,50 @@
+models:
+  llm:
+    # New Multi-Backend Config
+    qwen_1_5b:
+      backend: qwen
+      path: models/qwen2.5-1.5b-instruct
+      device: auto
+      dtype: bfloat16
+      max_context_length: 4096
+    
+    llama_8b:
+      backend: llama
+      path: models/llama-3.1-8b-instruct
+      device: auto
+      dtype: bfloat16
+      max_context_length: 8192
+
+    # Legacy fallback (needed if from_config is called directly without name)
+    hf_id: Qwen/Qwen2.5-1.5B-Instruct
+    local_path: models/qwen2.5-1.5b-instruct
+    dtype: bfloat16
+    device_map: auto
+
+  preference_extractor:
+    # Default/Legacy
+    default:
+    hf_id: Qwen/Qwen2.5-0.5B-Instruct
+    local_path: models/qwen2.5-0.5b-instruct
+    dtype: bfloat16
+    device_map: auto
+    # New SFT Extractor
+    qwen3_0_6b_sft:
+      path: saves/qwen3-0.6b-full-sft-h200/checkpoint-4358
+      prompt_template_path: fine_tuning_prompt_template.txt
+      device: auto
+      dtype: bfloat16
+      max_new_tokens: 512
+  embedding:
+    qwen3:
+      hf_id: Qwen/Qwen3-Embedding-8B
+      local_path: models/qwen3-embedding-8b
+    nemotron:
+      hf_id: nvidia/llama-embed-nemotron-8b
+      local_path: models/llama-embed-nemotron-8b
+  reranker:
+    qwen3_8b:
+      hf_id: Qwen/Qwen3-Reranker-8B
+      local_path: models/rerankers/qwen3-reranker-8b
+      dtype: bfloat16
+      device_map: auto
diff --git a/configs/qwen2.5_0.5b_full_sft.yaml b/configs/qwen2.5_0.5b_full_sft.yaml
new file mode 100644
index 0000000..ca1cca2
--- /dev/null
+++ b/configs/qwen2.5_0.5b_full_sft.yaml
@@ -0,0 +1,34 @@
+### Qwen2.5-0.5B Full SFT Config
+model_name_or_path: Qwen/Qwen2.5-0.5B-Instruct
+stage: sft
+do_train: true
+finetuning_type: full
+freeze_trainable_layers: 0
+
+dataset: preference_extractor_train
+template: qwen
+cutoff_len: 1024
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+output_dir: saves/qwen2.5-0.5b-full-sft
+logging_steps: 10
+save_strategy: steps
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+
+per_device_train_batch_size: 16
+gradient_accumulation_steps: 8
+learning_rate: 2.0e-5
+num_train_epochs: 1.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.05
+bf16: true
+flash_attn: fa2
+
+val_size: 0.01
+per_device_eval_batch_size: 16
+eval_strategy: steps
+eval_steps: 500
+
diff --git a/configs/qwen2.5_1.5b_full_sft.yaml b/configs/qwen2.5_1.5b_full_sft.yaml
new file mode 100644
index 0000000..e91176b
--- /dev/null
+++ b/configs/qwen2.5_1.5b_full_sft.yaml
@@ -0,0 +1,33 @@
+### Qwen2.5-1.5B Full SFT Config
+model_name_or_path: Qwen/Qwen2.5-1.5B-Instruct
+stage: sft
+do_train: true
+finetuning_type: full
+
+dataset: preference_extractor_train
+template: qwen
+cutoff_len: 1024
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+output_dir: saves/qwen2.5-1.5b-full-sft
+logging_steps: 10
+save_strategy: steps
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+
+per_device_train_batch_size: 8
+gradient_accumulation_steps: 16
+learning_rate: 2.0e-5
+num_train_epochs: 1.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.05
+bf16: true
+flash_attn: fa2
+
+val_size: 0.01
+per_device_eval_batch_size: 8
+eval_strategy: steps
+eval_steps: 500
+
diff --git a/configs/qwen3_0.6b_full_sft.yaml b/configs/qwen3_0.6b_full_sft.yaml
new file mode 100644
index 0000000..e41a419
--- /dev/null
+++ b/configs/qwen3_0.6b_full_sft.yaml
@@ -0,0 +1,35 @@
+### Qwen3-0.6B Full SFT Config (H200x4 Optimized)
+model_name_or_path: Qwen/Qwen3-0.6B
+stage: sft
+do_train: true
+finetuning_type: full
+freeze_trainable_layers: 0
+
+dataset: preference_extractor_train
+template: qwen
+cutoff_len: 1024
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+output_dir: saves/qwen3-0.6b-full-sft-h200
+logging_steps: 5
+save_strategy: steps
+save_steps: 200
+plot_loss: true
+overwrite_output_dir: true
+
+# H200x4 Configuration
+# Total Batch Size = 32 * 4 * 1 = 128
+per_device_train_batch_size: 32
+gradient_accumulation_steps: 1
+learning_rate: 2.0e-5
+num_train_epochs: 1.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.05
+bf16: true
+flash_attn: fa2
+
+val_size: 0.01
+per_device_eval_batch_size: 32
+eval_strategy: steps
+eval_steps: 200
diff --git a/configs/qwen3_1.7b_full_sft.yaml b/configs/qwen3_1.7b_full_sft.yaml
new file mode 100644
index 0000000..069c53d
--- /dev/null
+++ b/configs/qwen3_1.7b_full_sft.yaml
@@ -0,0 +1,34 @@
+### Qwen3-1.7B Full SFT Config (H200x4 Optimized)
+model_name_or_path: models/Qwen3-1.7B
+stage: sft
+do_train: true
+finetuning_type: full
+
+dataset: preference_extractor_train
+template: qwen
+cutoff_len: 1024
+overwrite_cache: true
+preprocessing_num_workers: 4
+
+output_dir: saves/qwen3-1.7b-full-sft-h200
+logging_steps: 5
+save_strategy: steps
+save_steps: 200
+plot_loss: true
+overwrite_output_dir: true
+
+# H200x4 Configuration
+# Total Batch Size = 32 * 4 * 1 = 128
+per_device_train_batch_size: 32
+gradient_accumulation_steps: 1
+learning_rate: 2.0e-5
+num_train_epochs: 1.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.05
+bf16: true
+flash_attn: fa2
+
+val_size: 0.01
+per_device_eval_batch_size: 32
+eval_strategy: steps
+eval_steps: 200
diff --git a/configs/reranker.yaml b/configs/reranker.yaml
new file mode 100644
index 0000000..c376fc7
--- /dev/null
+++ b/configs/reranker.yaml
@@ -0,0 +1,3 @@
+reranker:
+  default: qwen3_8b
+
diff --git a/configs/retrieval.yaml b/configs/retrieval.yaml
new file mode 100644
index 0000000..d2e100e
--- /dev/null
+++ b/configs/retrieval.yaml
@@ -0,0 +1,5 @@
+retrieval:
+  dense_topk: 64       # Initial recall count
+  rerank_topk: 8       # Count fed to LLM after rerank
+  pca_dim: 256
+
diff --git a/configs/user_model.yaml b/configs/user_model.yaml
new file mode 100644
index 0000000..7b8e230
--- /dev/null
+++ b/configs/user_model.yaml
@@ -0,0 +1,14 @@
+user_model:
+  item_dim: 256
+  user_dim: 256
+  beta_long: 0.1       # Enable personalization for Day 4
+  beta_short: 0.3
+  tau: 1.0
+  preference_extractor_name: qwen3_0_6b_sft # Switch to new extractor
+  rl:
+    eta_long: 1.0e-3
+    eta_short: 5.0e-3
+    ema_alpha: 0.05
+    short_decay: 0.1
+
+llm_name: llama_8b # Switch backend to Llama 3.1