Evidence Dial — models and adapters
Collection
RAW Tiny Aya grounding adapters: supervised AlignTune comparison and the Particle Sliders cohort. Non-commercial CC-BY-NC-4.0 base. • 5 items • Updated
Evidence Dial behavior adapter — expanded 138-context study
Root adapters now contain the unique DEV-selected G=0.00025, step 256, strength 1 raw checkpoint from fresh 138-context training. Both cohorts completed 1,024 updates; the latest checkpoint is not automatically the best. Custom Particle Sliders artifacts, not standard PEFT weights.
| Condition (32 fresh questions) | Mean behavior score |
|---|---|
| OFF | 0.686719 |
| Fixed prompt-only | 0.649375 |
| PG strength 0.5 — descriptive | 0.682734 |
| PG strength 1 — selected on DEV | 0.673906 |
| PG strength 2 — descriptive | 0.639062 |
The target is sympathetic, constructive support with independent judgment. The learned adapter receives the same ordinary user question, with no added behavior instruction. A separate fixed prompt-only condition measures what prompting already achieves. Scores are NanoGPT Jev model judgments, not accuracy or human preference rates.
Both fresh cohorts, using 138 contexts across 37 questions, finished 1,024 updates with raw weights and no EMA. Development selection across 24 settings chose G=0.00025, step 256, strength 1 (DEV 0.6950). Longer training did not improve the maximum. All 160 fresh-confirmation responses end naturally; the 4,096-token safety cap was not reached (longest: 575 tokens).
The selected adapter's fresh score is 0.673906 versus OFF 0.686719 and prompt-only 0.649375. ON minus OFF is −0.012812, with an individual 95% paired question-bootstrap interval [−0.047266, +0.020781]; the change from prompt-only is +0.024531 [−0.000547, +0.049687]. Neither establishes an advantage. The selected sympathy diagnostic decreases and sycophancy increases. Strength 2 regresses. Strength 0.5 is descriptive, not a new test-selected winner.
One synthetic corpus, one training seed, shared curation/evaluation judge and limited first-state supervision constrain generalization. The fresh questions were frozen before reading trained expanded-run development results; broader semantic families overlap. Length changes may affect the judgments; no length-matched benefit is established.
Load the native artifact using particle_sliders.inference.open_inference with evidence_dial.behavior_expanded_binding_v2.inference_backend(). Use the context's binding.response.__self__.model and slider.scaled(strength) with the unchanged user question. behavior_fullanswer_generate_v2.py records the actual greedy 4,096-token/native-EOS path. Install the pinned project code and CohereLabs/tiny-aya-global at revision af89d219b53ed9b13b8a4645f9c8028973510324; core 0.14.2/PG 0.9.5, full BF16/SDPA GPU backbone, no CPU offload. A fresh-machine replication has not been verified.
All DEV maximum ties are retained (one in this study). selection.json records raw-weight identities and the display rule fixed before confirmation. New versioned checkpoint files are under expanded138-h1024-v2/checkpoints. Earlier 37-pair native/raw weights remain in their historical checkpoint folder and previous model revision; old results are not re-labeled as full-answer confirmation.
A separate QA study tied SFT at 174/192 on DEV, while sealed 256-question results favored SFT (PG 221–229, SFT 230–234, OFF 142). No SFT parity or superiority claim.
- Thorough README and code
- Raw behavior adapter
- All states, predictions, logs and audits
- Model collection
- Dataset collection
Retain Tiny Aya CC-BY-NC-4.0 noncommercial and upstream code licenses.