How zyang1580/collm works
CoLLM is a research recommendation system that combines collaborative user/item representations with an LLM. Its principal workflow builds recommendation datasets and prompts, trains first on text-only inputs (LoRA tuning), then tunes the collaborative-information embedding (CIE) mapping with IDs, and evaluates through recommendation tasks and runners. The graph includes the documented Qwen variant and baseline training branch. Source excerpts show the MF/DIN and SASRec entrypoints and core orchestration; several model and dataset internals are indexed but not excerpted, so their detailed call relationships are omitted.
Main entry points: train_collm_mf_din.py CoLLM training · rec_base_task.py Recommendation task · minigpt4rec_v2.py CoLLM model · runner_base_rec.py Recommendation runner
Components 22
Training entrypoints
- CoLLM training
train_collm_mf_din.py - SASRec training
train_collm_sasrec.py - Baseline training
baseline_train_lightgcn_ood.py
Recommendation data
- Dataset builders
minigpt4/datasets/builders/rec_base_dataset_builder.py - Pair builder
minigpt4/datasets/builders/rec_pair_builder.py - Recommendation datasets
minigpt4/datasets/datasets/rec_datasets.py - GNN dataset
minigpt4/datasets/datasets/rec_gnndataset.py - Prompt processors
minigpt4/processors/rec_processors.py - Dataset utilities
minigpt4/datasets/data_utils.py
Recommendation models
- CoLLM model
minigpt4/models/minigpt4rec_v2.py - Qwen CoLLM
minigpt4/models/minigpt4rec_v2_qwen.py - Collaborative model
minigpt4/models/rec_model.py - Rec backbones
minigpt4/models/rec_base_models.py - LLM implementation
minigpt4/models/modeling_llama.py
Training runtime
- Recommendation task
minigpt4/tasks/rec_base_task.py - Task training loop
minigpt4/tasks/base_task.py - Recommendation runner
minigpt4/runners/runner_base_rec.py - Runner base
minigpt4/runners/runner_base.py - Configuration
minigpt4/common/config.py - Component registry
minigpt4/common/registry.py
Other components
- Researcher
- Pretrained LLM weights
Connections 27
- Researcher → CoLLM training: launches
README.md - CoLLM training → Configuration: loads config
train_collm_mf_din.py - CoLLM training → Recommendation task: builds task
train_collm_mf_din.py - CoLLM training → Recommendation runner: starts training
train_collm_mf_din.py - SASRec training → Recommendation task: builds task
train_collm_sasrec.py - SASRec training → Recommendation runner: starts training
train_collm_sasrec.py - Recommendation task → Task training loop: extends
minigpt4/tasks/rec_base_task.py - Recommendation task → CoLLM model: builds model
- CoLLM training → Dataset builders: loads datasets
- Dataset builders → Recommendation datasets: constructs
minigpt4/datasets/builders/rec_pair_builder.py - Dataset builders → Prompt processors: configures
- Recommendation datasets → Dataset utilities: uses utilities
- SASRec training → GNN dataset: loads dataset
train_collm_sasrec.py - Recommendation task → Dataset utilities: prepares samples
minigpt4/tasks/rec_base_task.py - Recommendation runner → Runner base: extends
minigpt4/runners/runner_base_rec.py - Recommendation runner → Recommendation task: evaluates with
minigpt4/runners/runner_base_rec.py - CoLLM training → Component registry: registers modules
train_collm_mf_din.py - Recommendation task → Component registry: selects components
minigpt4/tasks/rec_base_task.py - CoLLM model → Collaborative model: uses collaborative model
minigpt4/models/minigpt4rec_v2.py - Qwen CoLLM → Collaborative model: uses collaborative model
minigpt4/models/minigpt4rec_v2_qwen.py - CoLLM model → LLM implementation: uses LLM
minigpt4/models/minigpt4rec_v2.py - Qwen CoLLM → LLM implementation: uses LLM implementation
minigpt4/models/minigpt4rec_v2_qwen.py - Collaborative model → Rec backbones: uses backbones
minigpt4/models/rec_model.py - Pretrained LLM weights → CoLLM model: initializes
README.md - CoLLM training → CoLLM model: tunes LoRA first
README.md - CoLLM training → CoLLM model: then tunes CIE
README.md - Researcher → Baseline training: launches baseline