README.md

metadata

license: apache-2.0
tags:
  - diffusiongemma
  - gemma4
  - code-graphs
  - code-refactoring
  - latent-diffusion
  - ast
pretty_name: DiffusionGemma 2 Latent Code Graph Benchmarks & Raw Predictions
size_categories:
  - 1K<n<10K

DiffusionGemma 2: Latent Code-Graph Benchmarks & Raw Predictions

Standardized AST code-graph benchmark packs, continuous symbol embeddings, pre-registration (PREREG.json), linear evaluation probes, and raw .npy and .npz model prediction arrays for independent CPU rescoring of Latent Roadmaps: Soft-Token Re-Feeding in DiffusionGemma for Code Graphs.

• Author: bumblebeers (GitHub)
• License: Apache-2.0
• Code & Scoring Harness: bumblebeers/dgg2-latent-code-graphs
• Trained Two-Seed LoRA Adapters & Readout Heads: bumblebeer/dgg2-latent-code-graphs-adapters
• Standalone Libraries (v0.1.0): dgg-latent | dgg-codebook | dgg-vit-decoder
• Upstream transformers Multimodal Mask Fix: huggingface/transformers#50684

---

Repository Contents

1. Experiment A: 16x16 Visual Adjacency Grid Benchmark Packs (data/gridA_*)

• data/gridA_train1608.npz: 1,608 16-symbol Python AST subgraphs extracted across 3 training repositories (fastapi, requests, rich) with 22 fixed candidate states and 3 open-ended edit families.
• data/gridA_test360_httpx.npz: 360 held-out 16-symbol AST subgraphs from encode/httpx across seen, shuffled, unseen paraphrase, indirect intent, held-out single-pass, 2-step re-feeding, 3-step chain, and open-ended manifests.
• data/gridA_test360_click.npz: 360 held-out 16-symbol AST subgraphs from pallets/click for zero-shot cross-repository transfer.
• data/gridA_n32_*.npz & data/gridA_n48_*.npz: Multi-resolution 32x32 and 48x48 subgraph packs testing sub-patch packing limits (Gate A1).
• data/*.nodes.json: Per-subgraph AST symbol and line metadata.
• reader.npz (outputs/m1/reader.npz): Gate A1 closed-form Ridge probe mapping clean SigLIP latents h_v (256 x 1152) to the 8 discrete palette classes (1.000 clean cell accuracy across all 8 classes).

2. Experiment B: 256-Symbol Continuous Embedding Packs (data/nodeB_*)

• data/nodeB_train1608.json.gz: 1,608 256-symbol AST windows for single-site and multi-site symbol refactoring.
• data/nodeB_test360_httpx.json.gz & data/nodeB_test360_click.json.gz: Held-out httpx (N=360) and zero-shot click (N=360) evaluation packs (chance = 12.5%, mean-vector floor = 25.0%).
• data/nodeB_test360_httpx_unseenargs.json.gz & data/nodeB_test180_httpx_renameclass.json.gz: Unseen argument and class-rename splits.
• data/nodeB_multi_test360_httpx_k2.json.gz & k4.json.gz: 2-site and 4-site simultaneous and re-fed symbol-editing splits.
• emb.npz & emb_train.npz: Frozen sentence-transformers/all-MiniLM-L6-v2 (d=384) symbol embeddings and train-only standardization statistics.

3. Raw Model Prediction Arrays (predictions/ & outputs/m2/decoded_grids_all_arms.npz)

To enable full independent verification without requiring a 26B GPU, this dataset hosts both continuous latent .npy outputs and Ridge-decoded discrete (M, 16, 16) uint8 grid .npy/.npz files:

• predictions/expA_hv/<arm>/preds_*.npy: Raw continuous predicted SigLIP latent tensors (360 x 256 x 1152, float32) emitted by stock_generate_with_readout for two-seed LoRA DG (cg_lora_dg_seed42_ep3, cg_lora_dg_seed1337_ep3), cg_frozen_dg_ep3, and Non-DG baselines.
• predictions/expA_grids/<arm>/*.npy & outputs/m2/decoded_grids_all_arms.npz: Decoded (360, 16, 16) uint8 palette grids (0..7) across all evaluated arms — two-seed LoRA DG (httpx and two-seed zero-shot click), Frozen DG, two-seed Non-DG (Frozen Instr and Enc LoRA), and preliminary google/gemma-4-26B-A4B-it autoregressive references (2-shot gemma4_ar_prompted and 354-of-4,824-step single-seed gemma4_ar_lora_r64_s42 across all 4 primary splits).
• predictions/expB_vec/<arm>/*.npy: Raw standardized continuous symbol-vector predictions (M x 256 x 384, float32) for two-seed Experiment B LoRA DG (cg_expB_lora_dg_seed42_ep3, cg_expB_lora_dg_seed1337_ep3) and two-seed Non-DG baselines.

---

Independent CPU Rescoring Quickstart

git clone https://github.com/bumblebeers/dgg2-latent-code-graphs.git
cd dgg2-latent-code-graphs

# Rescore Experiment A directly from raw continuous h_v (.npy, shape 360x256x1152) + Ridge probe:
python3 cg/cg_score_preds.py \
  --pack data/gridA_test360_httpx.npz \
  --reader reader.npz \
  --manifest seen_bal360 \
  --pred predictions/expA_hv/cg_lora_dg_seed42_ep3/preds_cg_lora_dg_seed42_ep3_seen_bal360.npy

# Rescore decoded (360, 16, 16) uint8 palette grids (.npy), including Gemma 4 AR LoRA:
python3 cg/cg_score_preds.py \
  --pack data/gridA_test360_httpx.npz \
  --manifest refeed_heldout_comp_bal360 \
  --grids predictions/expA_grids/gemma4_ar_lora_r64_s42/refeed_heldout_comp_bal360.npy