Gemma-2-2B-IT Abliterated — BonfyreFPQ Native
Source: IlyaGusev/gemma-2-2b-it-abliterated
Format: BonfyreFPQ v4 seed programs (.fpq)
Compression: BF16 safetensors → FPQ seed programs
Compression Results
| Shard | Tensors | BF16 Source | FPQ Output | Ratio |
|---|---|---|---|---|
| model-part1.fpq | 270 | 4.60 GB | 279 MB | 16.9× |
| model-part2.fpq | 18 | 230 MB | 13 MB | 17.7× |
| Total | 288 | 4.83 GB | 292 MB | 16.9× |
- Avg bits/weight: 3.50 bpw (vs 16 bpw BF16)
- FP32 ratio: 9.1× (9972 MB → 292 MB on disk)
- Mode: COORD@3+QJL (3-bit Lloyd-Max coordinate quantization + Johnson-Lindenstrauss projection)
FPQ Format
Each .fpq file stores seed programs — combinator trees that regenerate weight tensors on demand — rather than the weights themselves. The format uses:
- COORD@3 mode: 3-bit Lloyd-Max scalar quantization per coordinate within E8 lattice blocks
- QJL residuals: Johnson-Lindenstrauss compressed residual bits per 256-dim block
- Shared base seeds: DAG-compressed common subexpressions across blocks
Requires bonfyre-fpq decompress to reconstruct weights for inference.
Aurekai / Bonfyre Native
Compressed with bonfyre-fpq compress from the Aurekai runtime:
bonfyre-fpq compress <shard.safetensors> <shard.fpq> --report
Inspect with:
bonfyre-fpq inspect model-part1.fpq
Source Attribution
This is a compressed redistribution of IlyaGusev/gemma-2-2b-it-abliterated, which is itself derived from google/gemma-2-2b-it. See source repositories for licensing.