OpenGisting: Qwen3-1.7B Gist vectors
This is a custom input-embedding adapter for an English-only customer-support/order-delivery demo with synthetic orders. It is not affiliated with Shopify. It needs the OpenGisting runtime and a separately downloaded, pinned Qwen3-1.7B base. It is not a standalone language model or a PEFT/AutoModel drop-in; no base weights, tokenizer files, or new visible vocabulary are packaged.
Artifact and compatibility
| Field | Value |
|---|---|
| Run | k16-v8-seed20261002 |
| Tensor | gist, float32, 16 × 2048; 32,768 values |
| File | gist.safetensors, 131,152 bytes |
| SHA-256 | a41e307cfe017300261c6e628d749bbf64922e6a0e34e880a5b6de8f5f3063a9 |
| Base revision | 70d244cc86ccca08cf5af4e1e306ecf908b1ad5e |
| Backend ID | transformers-mps-bfloat16 |
| Placeholder ID | 151669 |
Prompt assembly inserts the same placeholder ID 16 times. The runtime substitutes the 16 learned vectors at those input-embedding positions in order, without editing the base embedding table. Typing token spellings into chat does not activate Gist.
The tensor is stored as float32; the recorded runtime backend uses bfloat16. Preserve every manifest check: base revision/checksum metadata, rules, tools, dimensions, placeholder capacity, backend/dtype, and gist checksum. Dataset and training-configuration hashes also preserve provenance. Other backends need a matched artifact; do not bypass or rewrite the manifest to force loading.
The package contains the tensor, manifest.json, LICENSE, bilingual cards, and SHA256SUMS. The artifact and pinned base use Apache-2.0.
Download and use
Use source snapshot 758c021274bd4c971125da4435516467a9796d3d. Follow the dependency and pinned-base setup: Python 3.11, uv, the model extra, and the hf CLI. Run from the source repository root after setting GISTING_MODEL_DIR as described there.
hf download ImPanda/how-shopify-gisting-works --local-dir ./artifacts/gist/k16-v8
(cd ./artifacts/gist/k16-v8 && shasum -a 256 -c SHA256SUMS)
export GISTING_GIST_DIR=./artifacts/gist/k16-v8
Compare the tensor checksum with the SHA-256 above and in manifest.json; package checksums do not replace runtime compatibility checks. Follow the reviewed Full and Gist runtime commands. Keep GISTING_GIST_DIR pointing to the directory downloaded above.
Keep local synthetic transport and all runtime guards: order verification before data access, failure limits, cross-order isolation, consent checks, untrusted-text tokenization, fact checks, and public-trace redaction. The frozen tool set is lookup_order, handoff_to_human, and send_shipping_reminder. Learned vectors are not the security boundary; see architecture.
Historical training
Only the gist embedding was trained; teacher and student used the same frozen base weights. Recorded base-parameter hashes before and after training are identical.
The run used 92 train and 47 dev teacher examples, 138 optimizer steps, and chunk_mean initialization. Dev KL changed from 2.463852331204915 to 0.1581205797248642.
KL covers all retained response-token positions; there is no semantic decision/fact-only loss mask. Lower KL alone does not establish answer correctness.
The original teacher-data SHA-256 is recorded as dataset_sha256 in the manifest. That teacher artifact is not in this HF package. The source snapshot includes the recipe, templates, and sanitized data; see training.
Recorded evaluation and limits
Historical Full and Gist reports each record 933 cases. Read raw and final scores separately: runtime guards and renderers contribute to final outcomes, so these are not model-only successes.
In that frozen epoch, the fixed rules segment changes from 526 tokens to 19 (16 gist positions + 3 framing tokens); the tools segment remains 395 tokens. History and tool results remain ordinary inputs. See the measured boundary. Context distillation updates only the gist embedding; Qwen3-1.7B base weights remain fixed. Answer quality, latency, and transfer to new domains require separate evaluation. Changed rules/tools require a fresh matched artifact and evaluation; other languages and domains are outside this demo's scope.
Release preparation verified file/hash identity, current manifest compatibility, tensor key/dtype/shape/finite values, placeholder capacity, and report artifact identity. This was file/compatibility validation only: no new model generation or evaluation was performed for this release. Downloading the package and checking its hashes validates file transfer, not model execution.