Gemma Legal Grounded QLoRA v01 Corrective
Research/Educational Only
This experimental adapter is not production-ready, is not legal advice, and is unsuitable for
personalised legal advice, legal decisions, or other high-stakes use. It must be used with supplied context,
independent verification, and application-level refusal controls.
Provenance and training
- Base model:
google/gemma-2-2b-it
- Locked base revision:
299a8560bedf22ed1c72a8a11e7dce4a7f9f51f8
- Dataset: 1,498 human-reviewed grounded SFT records derived from Indian Contract Act provisions
- Method: QLoRA, NF4 double quantization, BF16 compute
- LoRA targets:
q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
- Adapter SHA-256:
126c4e26a8e2168ef339160f70c8d61be84cdb56893dea0dd34ffc16ae8af141
This adapter is the one-epoch corrective experiment trained on 24 human-approved refusal records and 16 unchanged grounded replay records at learning rate 1e-5.
Evaluation findings
The adapters show strong grounded extraction. In the post-correction held-out comparison, both the original
and corrective adapters passed 8/8 grounded/paraphrased cases with mean token F1 0.97097; there was no
grounded-performance regression. However, the corrective adapter failed 14 of 16 expanded integrity cases
covering insufficient context, personalised advice, false premises, and adversarial requests. The corrective
release gate therefore failed. This negative result is intentionally preserved as an experimental research finding.
Release evaluation SHA-256: 71fa5b93ce02d0cec87bd9dbe205c1cffc04d7574df6996878d6e17724d2a52a.
Intended use
Limited research and teaching on context-grounded extraction. A suitable host application must require context,
reject personalised or adversarial requests before inference, check whether context is sufficient, label every
response experimental, and avoid representing outputs as authoritative law.
Out-of-scope use
- Personalised legal advice or predictions
- Autonomous legal research, filing, compliance, or decision-making
- Answering without supplied context
- Production or safety-critical deployment
- Circumventing application refusal controls
Loading
Load the locked base revision in accordance with the upstream Gemma license, then attach this PEFT adapter.
The original tokenizer chat template is authoritative.
