gpodz-gee-0p8b-0.0002-20261010
2026-10-10. Latest build. Experimental, not served.
Gee 0.8B, version 0.0002, trained on gee identity pairs plus published gpodz company material. It is a retry of build gpodz-gee-0p8b-0.0001-20261010, which we revoked: Gee probes 13/15 (bar 15/15).
- Base model
- Qwen3.5-0.8B-Base, Apache-2.0
- Build
- gpodz-gee-0p8b-0.0002-20261010
- Released
- 2026-10-10
- Base revision
- Not recorded at download; record on next pull
- Fine-tune
- LoRA, rank 16
- License
- Apache-2.0, same as the base model
Training data
- G-1 narrow ruling: own published copy and Gee identity pairs only. Nothing published.
- Question and answer pairs from published site copy: 1860
- Identity pairs (says it is an AI, names no model, maker or cloud): 434
- Contrast pairs (customers, pricing, use cases, industries, trust): 262
- Refusal pairs: 74
- Split: 2367 train, 263 validation, 50 evaluation.
Training
- 1776 iterations over 3 epochs, batch 4, learning rate 1e-4.
- Final validation loss 0.26. Low validation loss on templated data does not carry over to held-out recall.
Test results
| Gate | Result | Bar | Outcome |
|---|---|---|---|
| Gee probes | 15/15 | 15/15 | Pass |
| Identity diversity (none naming a model, maker or cloud) | 10/10 | 6/10 | Pass |
| Leak suite | 0/500 | 0 leaks | Pass |
| Held-out accuracy | 9/50 | Informational for 0.8B | Informational, no pass bar at this size |
| Trust status questions | 1/15 | Informational | Informational, no pass bar at this size |
Every gate we ran on this build is listed, including the ones it does not pass. The comparison below shows earlier builds, including the one that failed identity diversity.
Builds compared
| Gate | Build 0000 | Build 0001 | Build 0002 |
|---|---|---|---|
| Gee probes (bar 15/15) | 15/15 pass | 13/15 FAIL | 15/15 pass |
| Accuracy (informational) | 14/50 | 12/50 | 9/50 |
| Trust status | 3/15 | 4/15 | 1/15 |
| Leak suite (bar 0) | 0/500 pass | 0/500 pass | 0/500 pass |
| Identity diversity (bar 6/10) | 4/10 FAIL | 6/10 pass | 10/10 pass |
| Train pairs / iterations | 731 / 549 | 731 / 549 | 2367 / 1776 |
| Validation loss (final) | 0.514 | 0.700 | 0.260 |
Not served. The live Gee bot (Gee 2026.10) answers with the hosted base model and its prompt; trained Gee builds are research only until a build passes every gate and is approved.
Known limits
- Accuracy fell to 9/50: held-out trust questions collapse to one blended answer.
- Low validation loss on templated data does not carry over to held-out recall.
- Identity spec file not yet written; base revision not recorded.
Use
Intended: Answer questions about Gpodz from published site copy. Say plainly that it is an AI and point to the right page.
Out of scope: Inventing customers, prices or certifications. Naming its base model, maker or hosting in chat. Anything not in published site copy.
Integrity
- Base weights (sha256)
- c2b1e5a17d9c1e27685d92ed9b382911ebb99955ecd89052d1721241adfbab6c
- Fine-tune (sha256)
- 4cf095736098d3e28a5d51839108060f5ea8c8340bcb2304e68677a12aae2b6b
- Training corpus (sha256)
- 017726fa5a10bc16ccd6a597e723a0b722c93d3e98eec6be4104156b398939b2