Repository navigation
Releases
Releases · oramasearch/base-agent-0.6b
Release v0.1.0
Released 2026-11-16. The first release. base-agent-0.6b packs the Orama base model, a fine-tune of Qwen3-0.6B, into a container image that the Orama operator copies onto every workspace Zetta's volume, so a native agent created without a model runs on it. Because it is the first release, this page lists everything in it.
What it includes
The image: oramasearch/base-agent-0.6b:v0.1.0
- The checkpoint
v16of the Hugging Face modelOramaSearch/base-agent-0.6b, at Hub revision3f6eb0850f3c9ed1743c56b9aff80ba8ddb73084, under/model/:config.json,generation_config.json,model.safetensors(fp32, 2,384,234,937 bytes),tokenizer.json,tokenizer_config.json,vocab.json,merges.txtand the version'sREADME.md. /bin/shfrom BusyBox 1.37, the base image, pinned by its multi-platform index digest (sha256:a7dbc319…), because the operator's copy script runs in this image (#5).- Every file readable by uid 10001, the Zetta node's user; the image runs as that user.
- A build-time check: the build fails unless
/modelholds the config, the weights and a tokenizer the engine can load. - Built for
linux/amd64, from the tag, with BuildKit: digestsha256:0950dcfed899bb44d199dcfbb1a851dfcacd3d78009327368a6272898829d910.
The repository
- The
Dockerfileand the contract inREADME.md: what/modelholds, the shell, the user, and the rule that a published tag is never moved. - The build recipe: a context directory with one folder,
checkpoint/, taken from a local checkpoint or downloaded from the Hub. VERSIONholds the release version;tools/test-version.pychecks that these notes exist, are indexed and are named in the README.
Licence and project files
- The repository's files are MIT:
LICENSE,COPYRIGHT, anSPDX-License-Identifier: MITline in every source file,CONTRIBUTING.md,CLA.md(a copyright-assignment contributor licence agreement) andCODE_OF_CONDUCT.md(#7). - The weights in the image are under the Apache License 2.0. They are a modified version of Qwen3-0.6B (Apache-2.0), further trained on Orama's own synthetic corpus, and the modified weights are released under the same licence.
NOTICEsays so, as the Apache License 2.0 (section 4) requires;THIRD_PARTY_LICENSEScarries the Qwen3-0.6B licence and BusyBox's (#7). SECURITY.md: how to report a vulnerability privately (#3).- The public tree carries nothing internal (#8).
- A software bill of materials,
sbom.spdx.jsonandsbom.cdx.json, regenerated bytools/sbom.py(#7).
Changes
| Change | What |
|---|---|
| first commit | the Dockerfile and README: the base model checkpoint as a model image |
| #3 | SECURITY.md |
| #5 | the BusyBox base pinned by digest |
| #8 | the tree prepared for the public export |
| #7 | MIT licence, SPDX headers, NOTICE, third-party notices, contributor documents and the SBOM |
| this release | VERSION, tools/test-version.py, these notes, and the README's release line |
What it does not include
- No
v16image tag. The operator's default base model image namesoramasearch/base-agent-0.6b:v16, which is not pushed by this release (#6). The release image is taggedv0.1.0. - No bf16 image. The checkpoint is fp32 (about 2.4 GB); a bf16 copy would halve the image and the volume cost.
- Checkpoint pinning and checksums in the build are not included: #4 (pin the Hub revision and verify the files' checksums) is deferred, because it changes the build's inputs. This release records the revision and the weights' checksum below instead.
- No
linux/arm64image.
Known issues
- BusyBox (GPL-2.0) is the image's base (#10). The image carries BusyBox unmodified, and its source is at https://busybox.net/downloads/busybox-1.37.0.tar.bz2. Whether a written source offer must accompany the image, or the image moves to
scratchwith static binaries, is open. - The training corpus (#9): the weights were fine-tuned on a synthetic corpus partly produced with a hosted model, and the corpus and the Hugging Face model card are still to be cleaned before they are published. The Hugging Face repository
OramaSearch/base-agent-0.6bis not public yet. - The legal review of the weights' licence decided in #7 is open.
- The checkpoint is a bootstrap, not a production model. The
v16folder's own README, in the image at/model/README.md, says so: it can fabricate claims and confuse a clarifying question with an action. It is the starting point for reinforcement training.
Verification
| Proof | Result |
|---|---|
python3 tools/test-version.py |
passes |
Image build from the tag on BuildKit (linux/amd64), including the Dockerfile's checkpoint check |
passes; oramasearch/base-agent-0.6b:v0.1.0 and :main-72d363e name sha256:0950dcfed899bb44d199dcfbb1a851dfcacd3d78009327368a6272898829d910 (2.2 GB compressed) |
model.safetensors sha256 against the Hub's LFS record |
4f9c4dc7d0c8f3bd6707bd635b4d14d207a62a406ede35841c3ae2a780674819, matches |
tokenizer.json sha256 against the Hub's LFS record |
aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4, matches |
tools/spdx-headers.py --check |
passes |
tools/sbom.py --check |
passes, regenerated after the last change |
Not run: loading the image's checkpoint in a Zetta node, and the operator's copy into a workspace volume. The release is not deployed to any cluster.