Skip to content

weights.md

Latest commit

 

History

History
102 lines (71 loc) · 3.83 KB

weights.md

File metadata and controls

102 lines (71 loc) · 3.83 KB
Raw

Weights

PathoSynVLM has code in Git and model artifacts outside Git.

Official model repository: AtlasAnalyticsLab/PathoSynVLM.

The intended user experience is:

  1. The trained Stage 2 model is exported once into the inference layout below.
  2. The exported directory is hosted at AtlasAnalyticsLab/PathoSynVLM.
  3. Users download the exported directory into $PATHOSYNVLM_WEIGHTS_ROOT/pathosynvlm-stage2-main/.
  4. Users run inference or evaluation directly from those weights.

Users should not need to train the model just to generate a report, as long as the released weights are available.

What Users Download

The weight package should have this layout:

$PATHOSYNVLM_WEIGHTS_ROOT/pathosynvlm-stage2-main/
  config.json
  vlm_state.pt
  tokenizer/
  llm/
  best_checkpoint_summary.json

llm/ is a merged Hugging Face model directory for the paper release. This matters because the paper run used unfreeze_llm_base=true; a LoRA adapter alone is not sufficient for exact reruns unless the base updates are also included.

Download the official weights with:

export PATHOSYNVLM_HF_REPO=AtlasAnalyticsLab/PathoSynVLM
source configs/paths.example.env

hf download "$PATHOSYNVLM_HF_REPO" \
  --local-dir "$PATHOSYNVLM_WEIGHTS_ROOT/pathosynvlm-stage2-main"

The prepared Hugging Face upload root is the model repository itself: it contains the real llm/model.safetensors and vlm_state.pt files, not symlinks or pointer files. See docs/huggingface_release.md for the exact upload-root validation steps.

What export_release_weights.py Does

scripts/export_release_weights.py does not download weights from the internet and does not recreate the paper weights by itself. It converts a completed local training run into a clean inference package.

Use it when:

  • You are the author preparing the official model release.
  • You ran the training pipeline and want to package your own checkpoint.

Do not use it when:

  • You only want to run the pretrained model. In that case, download the released package instead.

Export command for retraining or packaging a new checkpoint:

python scripts/export_release_weights.py \
  --run_dir "$PATHOSYNVLM_RUNS_ROOT/stage2_main" \
  --output_dir "$PATHOSYNVLM_WEIGHTS_ROOT/pathosynvlm-stage2-main" \
  --overwrite

This script:

  1. Reads train_args.json and the selected trainer checkpoint.
  2. Loads the base Qwen2.5-3B-Instruct model.
  3. Applies LoRA if the run used LoRA.
  4. Loads full LLM state if present in trainer_state_step_<N>.pt.
  5. Merges LoRA into the LLM unless --no_merge_lora is set.
  6. Saves llm/, tokenizer/, vlm_state.pt, and config.json.

The raw trainer state can be large because it may include optimizer and full LLM state. Run export on a compute node with enough CPU RAM/GPU memory. The resulting config.json is sanitized to avoid embedding local absolute training paths.

Inference

python scripts/generate_case_report.py \
  --embeddings HISTAI-skin-b2/conch_v15/5x_512/patches/example_1.h5 \
               HISTAI-skin-b2/conch_v15/5x_512/patches/example_2.h5 \
  --output_json report.json

The default --weights path is $PATHOSYNVLM_WEIGHTS_ROOT/pathosynvlm-stage2-main, and relative --embeddings values are resolved under $PATHOSYNVLM_EMBEDDINGS_ROOT. You can also pass absolute file paths.

The output is expected to follow:

Diagnosis: ...
Certainty: ...
Conclusion: ...

Usage Modes

Goal Need released weights? Need to train?
Generate a report for a case yes no
Evaluate the official model on prepared HISTAI embeddings yes no
Run the full training pipeline no yes
Create your own redistributable model package no yes, then export