README.md

---
pretty_name: bloomery reference sets
license: other
tags:
  - bloomery
  - test-fixtures
---

# bloomery reference sets

The oracle files of the fixture tier of [bloomery](https://github.com/midagedev/bloomery): intermediate tensors
that ik_llama.cpp dumped while it ran bloomery's fixture models (the image tower's taps: mainline llama.cpp's mtmd). A fixture is a GGUF with random weights and the
shape of a real model. The sets are raw little-endian f32 tensors plus a `MANIFEST.tsv` each. They hold no model
weights. bloomery's gates compare the engine against them.

Sets: 18. Files: 29978. Size: 7.60 GB.

| family | fixture dir | sets | files | size |
|---|---|---|---|---|
| `deepseek41` | `/models/fixtures/v41` | 4 | 5507 | 1.02 GB |
| `glm5next` | `/models/fixtures/glm5next` | 8 | 10458 | 4.91 GB |
| `qwen4exp` | `/models/fixtures/qwen38` | 6 | 14013 | 1.67 GB |

## Make the fixture files

The sets belong to fixtures, and the fixtures are not in this repo. `fixture fetch` makes one
from the headers of its source model's files (a few MB per shard); its table of sources is in the bloomery repo
(`crates/model/src/bin/fixture.rs`, `SPECS`):

```
fixture fetch deepseek41 /models/fixtures/v41
fixture fetch glm5next /models/fixtures/glm5next
fixture fetch qwen4exp /models/fixtures/qwen38
```

The DeepSeek V4.1 DSpark draft fixture cannot be fetched, so no set of it is here. The set
`ref-mtp/fx_qwen4exp_prose64_n64_k1` also names the fixture's MTP draft file,
`/models/fixtures/qwen38/mtp-Qwen3.8-Flash-Next-shared-Q8_0.gguf`, which `fixture fetch` makes with the fixture.

## The path is part of the identity

A set states the fixture it was dumped from by its full path, `/models/fixtures/<dir>/<file>`. Make the fixtures at
those paths. On a machine without `/models`, make `/models/fixtures` a symlink to where you keep them.

## Put the sets in place

The repo mirrors `$BLOOMERY_DATA`: a set is `fx_.../<file>` or `ref-mtp/fx_.../<file>`, the path the gates read. One
download is the whole install:

```
hf download --repo-type dataset midagedev/bloomery-refsets --local-dir "$BLOOMERY_DATA"
```

To take one family, name its sets with `--include`. The sets of each family:

- `deepseek41`: `fx_ref_deepseek41_d1_every_node`, `fx_ref_deepseek41_d1n_every_node`, `fx_ref_deepseek41_d2_every_node`, `fx_ref_deepseek41_step4_every_node`
- `glm5next`: `fx_ref_glm5next`, `fx_ref_glm5next_d16kdsa`, `fx_ref_glm5next_d1k`, `fx_ref_glm5next_d3kdsa`, `fx_ref_glm5next_step4`, `fx_ref_glm5next_step4_every_node`, `ref-mtp/fx_prose64_n64_k1`
- `qwen4exp`: `fx_ref_qwen4exp`, `fx_ref_qwen4exp_d1k`, `fx_ref_qwen4exp_d3k`, `fx_ref_qwen4exp_step4`, `fx_ref_qwen4exp_step4_every_node`, `ref-mtp/fx_qwen4exp_prose64_n64_k1`

Optional check, from `$BLOOMERY_DATA`: `sha256sum --quiet -c SHA256SUMS`. It lists every file of the repo, so on a
partial download it reports the files you did not take as missing. The download also puts `README.md` and
`SHA256SUMS` in `$BLOOMERY_DATA`.

## Licences

Each fixture takes its shape from the header of one source model, and each set holds outputs of an engine that ran
that fixture. The source models' licences are below, and each text is in `LICENCES/`. Read them before you reuse a set
outside bloomery's gates.

- DeepSeek V4.1 Flash (family `deepseek41`): MIT, [`LICENCES/deepseek41.txt`](LICENCES/deepseek41.txt), a copy of [the base model's LICENSE](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/2cba9e42aa026125f3ed06c6d98c1db82f7ca027/LICENSE)
- GLM-5.3-Flash (family `glm5next`): MIT, [`LICENCES/glm5next.txt`](LICENCES/glm5next.txt), a copy of [the base model's LICENSE](https://huggingface.co/zai-org/GLM-5.3-Flash/blob/eb9eb208eb0d988989d07a6a12d0fdeb5f52574a/LICENSE)
- Qwen3.8-Flash-Next (family `qwen4exp`): Qwen Community License 1.0, [`LICENCES/qwen4exp.txt`](LICENCES/qwen4exp.txt), a copy of [the base model's LICENSE](https://huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/de4b8e4d43b917e7706784d8bb445c9af86a3540/LICENSE)

- ik_llama.cpp, the engine that dumped the sets: MIT, [`LICENCES/ik_llama.cpp.txt`](LICENCES/ik_llama.cpp.txt), a copy of [its LICENSE](https://github.com/ikawrakow/ik_llama.cpp/blob/c10fbbcccd69cb1c66ddeb39601b0cd71701f7c3/LICENSE)

`fx_ref_glm5next_vis_taps` holds the GLM-5.3-Flash image tower's taps (`vit_out`, `after_fc_proj`, `after_ffn_proj`
and the final embeddings of nine images). Mainline llama.cpp's mtmd dumped it at build 36a73916e, on the fixture's
text model respelled in mainline's names; its `# respell` line names the fixture file it was cut from.

The fixture files are not here and hold no weights of the source models.

## Prompt corpora

Token ids, one per line, of the prose prompts bloomery's runs feed.

| file | ids | sha256 |
|---|---|---|
| `glm5next/corpus-prose.ids` | 71727 | `8af07981c1749170b57d424cff5274b89be063c1eb44ffa3a443ddc16642bb64` |
| `engram/corpus-prose.ids` | 75268 | `9c93ed03cfb513ab36520492de19330cfe779f1505ce07647942a60a9dd12cb2` |

Put them at `$BLOOMERY_DATA/glm5next/corpus-prose.ids` and `$BLOOMERY_DATA/engram/corpus-prose.ids`.

## Left out

Not published, because they cannot be remade off the maintainers' machine byte for byte, or their source is not
verified:

- `fx_iq4xs_ref_glm5next`: a set of the fixture variant 'iq4xs': the variant file is requantized on the box and cannot be made again byte for byte off it
- `fx_iq4xs_ref_glm5next_step4`: a set of the fixture variant 'iq4xs': the variant file is requantized on the box and cannot be made again byte for byte off it
- `fx_iq4xs_ref_glm5next_step4_every_node`: a set of the fixture variant 'iq4xs': the variant file is requantized on the box and cannot be made again byte for byte off it
- `ref-mtp/fx_iq4xs_prose64_n64_k1`: a set of the fixture variant 'iq4xs': the variant file is requantized on the box and cannot be made again byte for byte off it