diff options
| author | YurenHao0426 <Blackhao0426@gmail.com> | 2026-03-18 18:25:09 -0500 |
|---|---|---|
| committer | YurenHao0426 <Blackhao0426@gmail.com> | 2026-03-18 18:25:09 -0500 |
| commit | b6c3e4e51eeab703b40284459c6e9fff2151216c (patch) | |
| tree | 221410886f23214575f93b9ef44fa8431c9a6dfc /configs | |
Initial release: VARS - personalized LLM with RAG and user vector learning
Diffstat (limited to 'configs')
| -rw-r--r-- | configs/base.yaml | 0 | ||||
| -rw-r--r-- | configs/local_models.yaml | 66 | ||||
| -rw-r--r-- | configs/reranker.yaml | 3 | ||||
| -rw-r--r-- | configs/retrieval.yaml | 5 | ||||
| -rw-r--r-- | configs/user_model.yaml | 14 |
5 files changed, 88 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..ea001ea --- /dev/null +++ b/configs/local_models.yaml @@ -0,0 +1,66 @@ +# Base path for all models +_base_path: &base ./ + +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 + + # vLLM backend for high-throughput experiments + llama_8b_vllm: + backend: vllm + path: .//models/llama-3.1-8b-instruct + vllm_url: http://localhost:8003/v1 + model_name: meta-llama/Llama-3.1-8B-Instruct + 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: .//models/pref-extractor-qwen3-0.6b-sft + 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 + bge_base: + hf_id: BAAI/bge-reranker-base + local_path: .//models/rerankers/bge-reranker-base + dtype: float16 + device_map: auto 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 |