code/implementation2.md

Spectra-Reason-GCD Implementation

This document details the implementation of the Spectra-Reason-GCD pipeline, which integrates a spectral encoder (DreaMS + Perceiver Resampler), an LLM reasoning engine (Llama-3-8B with LoRA), and grammar-constrained decoding (GCD) for valid SMILES generation.

1. Environment Setup

Use the provided specrag conda environment.

source /cluster/tufts/liulab/lib/anaconda3/etc/profile.d/conda.sh
conda activate specrag

2. Directory Structure

Spec-RAG/
├── data/                   # Dataset JSONL files
├── grammars/
│   └── smiles.ebnf         # SMILES grammar for GCD
├── scripts/
│   ├── train_stage1_bridge.py
│   ├── train_stage2_cot_lora.py
│   ├── generate_gcd.py
│   ├── create_test_spectra.py
│   └── ...
├── spec_rag/               # Python package
│   ├── cot_system2.py      # Chain-of-Thought logic
│   ├── perceiver_resampler.py
│   ├── spectra_reason_encoder.py
│   ├── gcd_inference.py
│   └── ...
└── out/                    # Checkpoints and logs

3. Training Pipeline

Stage 1: Encoder Pre-training (Optional but Recommended)

Trains the Perceiver Resampler to bridge DreaMS embeddings to the LLM's embedding space.

accelerate launch scripts/train_stage1_bridge.py \
    --train-jsonl data/train.jsonl \
    --output-dir out/stage1 \
    --dreams-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/data/ssl_model.ckpt \
    --llm-name meta-llama/Meta-Llama-3-8B-Instruct \
    --epochs 5 \
    --batch-size 8 \
    --mixed_precision bf16

Stage 2: End-to-End CoT Fine-tuning

Fine-tunes the LLM (LoRA) and Perceiver Resampler jointly using Chain-of-Thought data.

accelerate launch scripts/train_stage2_cot_lora.py \
    --train-jsonl data/train_cot.jsonl \
    --val-jsonl data/val_cot.jsonl \
    --output-dir out/stage2 \
    --dreams-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/data/ssl_model.ckpt \
    --stage1-perceiver out/stage1/perceiver.pt \
    --llm-name meta-llama/Meta-Llama-3-8B-Instruct \
    --epochs 3 \
    --batch-size 4 \
    --mixed_precision bf16

Note: Monitor the training loss. It should decrease significantly. If loss remains high or constant, check that the dreams-ckpt is loading correctly and data is valid.

4. Inference

Run inference on test spectra. We use Grammar-Constrained Decoding (GCD) to ensure valid SMILES output.

Important:

1. Ensure --encoder-dir points to out/stage2 to load the trained Perceiver.
2. Use --no-gcd first to debug raw model output if you suspect issues.
3. If output is garbage (e.g., repeating 0.0.0...), check training loss and DreaMS loading.

python scripts/generate_gcd.py \
    --input-jsonl data/test_spectra.jsonl \
    --output-jsonl out/predictions.jsonl \
    --model-dir out/stage2 \
    --encoder-dir out/stage2 \
    --dreams-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/data/ssl_model.ckpt \
    --grammar grammars/smiles.ebnf \
    --max-peaks 60 \
    --max-new-tokens 200 \
    --no-gcd  # Remove this flag to enable Grammar Constraint

5. Troubleshooting

• Garbage Output ("0.0.0.0...", " ..."):
  • This indicates the model is in a degenerate state.
  • Check Training: Ensure training loss decreased. If loss was ~0.0 or constant, training failed.
  • Check Input: Verify test_spectra.jsonl contains valid peaks (m/z, intensity).
  • Check DreaMS: Ensure dreams-ckpt path is correct and accessible.
• Empty Output:
  • Can happen if max_new_tokens is too small or EOS token is generated immediately.
  • The inference script has been patched to correctly handle inputs_embeds generation output.
• "No such file or directory":
  • Ensure you run scripts from the Spec-RAG root directory.
  • Verify data/ folder contains required JSONL files.

6. Evaluation

(Use chemprop or rdkit based scripts to evaluate validity, novelty, and exact match accuracy of out/predictions.jsonl).