Pyrodit RM · R7 recovery study
Risk detection: training continuity and contextual precision
The R7 study evaluated full-model training continuity and broader data coverage against predefined acceptance criteria. It clarified the trade-off between risk detection and contextual false alarms, while retaining R5 as the serving baseline. The findings inform the next controlled comparisons of language pretraining and label consistency.
R7 evaluation complete · R5 retained as the serving baseline
Research question
Can retaining the full model and restoring broader training data improve detection without increasing false alarms or losing existing capabilities?
Method
In the study dated 24 September 2026, two candidates continued training from the full R3 model on 42,969 examples, each for two complete passes. Acceptance criteria were fixed before training. A separate word-pattern model used the same examples with observed toxicity labels to assess the signal available in the data.
Results
At a 5% false-alarm rate, the two final R7 candidates detected 49.33% and 51.33% of risky content in 750 CivilComments development examples, against a 90% target. Across 44 saved training states, no candidate met the complete set of acceptance criteria for both detection and preservation of existing capabilities. R5 and its existing input profile were retained; the study did not introduce a replacement serving model.
Training continuity and selection controls
The audit identified how encoder and decision-layer initialization, training-data coverage and checkpoint selection affected the experimental pipeline. R7 addressed these mechanisms and preserved independently reviewable records. Their individual contribution to model performance was not isolated in this study.
Contextual precision in the control comparison
The word-pattern control detected 125 of 150 risky CivilComments examples, with 29 false alarms among 600 harmless examples. At the same threshold, it flagged 154 of 847 harmless context examples. This illustrates why detection strength must be assessed together with context-sensitive precision. The control remained below the 90% detection target, did not train personal-attack detection and was not deployed.
Next controlled comparisons
The study directs subsequent comparisons toward language pretraining and differences in how source datasets label harmful content. Additional pretraining has not yet been evaluated. Reproducible training records support comparison; serving acceptance still requires the complete quality criteria.
Limitations
These are English development results, with a separately selected threshold for each evaluation group, not one validated production threshold. They do not establish performance in Turkish, German or other risk categories. The context set had been examined before and is not an untouched independent test. No final context test was run for R7 because no candidate qualified; the reported context figures belong to the separate word-pattern control.
About this report
These are technical research reports, not peer-reviewed publications. Each report identifies its source, measurement conditions and limitations. The evidence section explains what can be downloaded or reproduced.
Technical notes and source records
Source: risk-recovery-20260924. R7: learning rates 1e-5 and 3e-5, two epochs each, 12,231,040 tokens and 10,744 optimizer steps in total. All 44 checkpoints were independently checked; none qualified. Final Civil ROC-AUC: 0.873811 / 0.877656; AP: 0.687405 / 0.680610. The same-data lexical control used 41,335 observed toxicity labels, excluding 1,634 unknown labels. Civil AUC: 0.954533; AP: 0.893655; context recall: 94.68%, FPR: 18.18%. Recall measures detected risks; FPR measures false alarms among harmless inputs. AUC and AP summarise score ranking. The control learns toxicity only, while R7 also learns harassment; this is not a single-variable architecture comparison.
The file names and SHA-256 fingerprints identify the exact source versions used to prepare this report.
- risk-recovery-20260924/README.md
SHA-256: f34dfa887481040dbf14f4034aca218c640e9b6a63df1e8ff18197721c4d8817 - risk-recovery-20260924/r7/summary.json
SHA-256: a91e57ecc1543b601563661d3219356f469c13cacd876d78a144bc4cdd1442ba - risk-recovery-20260924/full-replay-lexical-control/README.md
SHA-256: 8cac0bf35fbb55bc7ef7c7e38fc0095c91db6e9f235f65f8039716c0bfb73873 - risk-recovery-20260924/full-replay-lexical-control/summary.json
SHA-256: 0a6633cfb73f00874eb16b90ee9262dfe77b754c3d5198fd937817a1251bd4dd