Skip to content
EdgeLex

AI & Intelligence · Training Center

Frontier, local open-source — or your firm's own model, trained on your own work.

Included with EdgeLex · Firm-model training on your infrastructure

app.edgelex.com — Lex Training CenterMC

🎓 Lex Training Center

Manage feedback, training examples, and track Lex quality.

OverviewFeedback ExplorerDatasetsTraining ExamplesAnalyticsTraces

Total Feedback

8

Thumbs Up

3

Thumbs Down

5

With Corrections

5

Runtime few-shotTrainable SFTEval holdoutTool failure skipLegal safety reviewPrivacy reviewDeterministic renderer
Clean-Baseline Training Factory Targetfirm_model_v2_clean_baseline_001 ▾

Legacy feedback remains review evidence. Dataset growth must use selected manual examples or selected traces.

Status
draft
Lineage
firm_model_v2
Records
0
Created
6/30/2026
Pending Reviews0 pending

✅ All clear — no examples need review.

The problem

Every legal AI vendor picks your model for you — even “model choice” usually means choosing which cloud. Owning the intelligence layer should include the option to own the model itself.

How a firm model actually gets better

Captured. Curated. Assembled. Measured. Traced.

Follow the path from a thumbs-down to a better firm model — every screen below is the real product.

01 · Captured

Feedback is signal, never training data.

Thumbs, corrections, rephrases — captured from real work and kept as review evidence. The banner says it plainly: feedback rows are not imported into the clean baseline. A correction becomes training data only after a human recreates it as a curated example.

Feedback Explorer — signal review only
OverviewFeedback ExplorerDatasetsTraining ExamplesAnalyticsTraces
Feedback Explorer is signal review only. Feedback rows are not directly imported into the clean baseline. Recreate trusted cases as manual examples, or use selected completed traces from the Traces tab.

👍 Evalsignal only

Q: “now, update your answer”

👍 Legalsignal only

Q: “explain from a lawyer's perspective how the court rules work”

👎 SFTsignal only

Q: “set a reminder to call the client in the morning”

👎 promoted SFT pre-baseline

Q: “Hi”

Lex Response

Hello! How can I assist you today?

User Correction

Always start the response with my name, which is “Maya”.

LLM Curation: PROMOTED · fallback — substantial correction auto-promoted for review

02 · Curated

Examples are edited by hand, with a quality bar.

Each candidate example carries its category, its dataset route — few-shot, eval, SFT — and a quality score. A rephrased question gets flagged NEEDS HUMAN EDIT and sits at 40% until someone fixes it. Import is a deliberate act, not a pipeline.

Training Examples — curated by hand, imported deliberately
OverviewFeedback ExplorerDatasetsTraining ExamplesAnalyticsTraces
All status ▾All categories ▾4 examples🎓 Import Selected+ Create Example
QuestionDatasetQualityStatus
show all meetings for today[NEEDS HUMAN EDIT] User rephrased as: “Do I have any meetings today?” — the orig…calendar_scheduling · rephraseReview40%● review
HiMaya, hello. How can I assist you today?…GeneralEval70%
Summarize the selected document.# Summary of Selected Document ## Document Version — **Document summarized:** *…Summary (legacy)Few-shot70%
Summarize the selected document.whenever you summarize a document, and if that document has more than one versio…GeneralEval70%

03 · Assembled

The dataset refuses to ship until it's honest.

Readiness fails closed: empty train split, missing eval holdout — export blocked, dry-run not ready. The holdout exists so the firm can measure improvement on examples the model never saw.

Datasets — readiness fails closed
OverviewFeedback ExplorerDatasetsTraining ExamplesAnalyticsTraces
firm_model_v2_clean_baseline_001 · draft ▾⟳ Run Readiness⬇ Export Training JSONLdraftfirm_model_v2Qwen3-14B base
ReadinessBlocked

JSONL export: blocked · Training dry-run: not ready

Blockers

train_split_emptyeval_holdout_missing

Warnings

validation_split_empty

Splits

None

Sources

None

Training routes

None

Safety labels

None

04 · Measured

Quality is a trend line, not a vibe.

Category breakdown shows where the feedback concentrates; the weekly trend shows positive-signal share climbing — always measured against the eval holdout, never the training set.

Analytics — where Lex is getting better
OverviewFeedback ExplorerDatasetsTraining ExamplesAnalyticsTraces
Category Breakdown

Feedback volume by category · last 30 days

Summary & review34%

Court rules & deadlines22%

General assistance17%

Calendar & scheduling15%

Drafting12%

Weekly Quality Trend

Positive-signal share of rated turns

58%Jul 6
61%Jul 13
60%Jul 20
66%Jul 27
69%Aug 3
71%Aug 10

▲ +13 pts since curation began — measured against the eval holdout, never the training set.

05 · Traced

Every run is evidence.

The edgeflow trace log records each run's intent, model, tokens, and latency — deterministic steps included. Completed traces can be selected and imported as training candidates, so the dataset grows from work that actually happened.

Traces — every run inspectable, importable as training evidence
OverviewFeedback ExplorerDatasetsTraining ExamplesAnalyticsTraces
AllCompletedErroredPartial☐ With feedback onlyfirm_model_v2_clean_baseline_001 ▾🎓 Import Selected100 runs
StartedGraphIntentModelTokensLatencyStatus
Aug 15, 10:17 AMedgeflowdocument_fact_extractionSummarize the selected document.Claude Sonnet 54,40769.6scompleted
Aug 15, 10:13 AMedgeflowcourt_deadline_computationUsing the selected Matter, determine the last permissible date…— (deterministic)14.3scompleted
Aug 15, 10:07 AMedgeflowisolated_graph_nodeYou are executing one isolated, read-only node of a registered…Claude Sonnet 51,20420.8scompleted
Aug 15, 9:58 AMedgeflowdocument_fact_extractionSummarize the selected document.GPT-4.13,88248.1scompleted
Aug 15, 9:41 AMedgeflowgeneral_chatshow active mattersQwen3 14B · local9626.2scompleted

The difference

Nothing auto-trains. Ever.

The lazy version of “your model learns from your firm” is a pipeline that trains on whatever lawyers type. EdgeLex refuses that: feedback is signal, examples are curated by hand, safety and privacy review gate every record, the dataset blocks itself until the holdout is honest — and the weights never leave your deployment.

Where the model runs →

Signal-only feedback

Thumbs and corrections inform curation; they never flow into training directly.

Fail-closed datasets

Missing holdout or empty split blocks export — readiness is enforced, not suggested.

Holdout honesty

Improvement is measured on examples the model never saw. Anything else is theater.

Capabilities

What it does

Any model, one governance layer

Frontier models with your own keys, local open-source models on your hardware, or your firm's own fine-tuned model — all governed by the same default-deny policy, capability binding, and mutation-safety routing.

Feedback from real work

Ratings, inline corrections, rephrase detection, topic abandons, and natural follow-ups are captured where lawyers actually work — a living record of where Lex helped and where it fell short.

Signal, not training data

Feedback is signal-only review evidence — nothing auto-trains. Dataset growth uses lawyer-recreated manual examples or explicitly selected, completed traces. A thumbs-down never silently rewires the model.

The Clean-Baseline Training Factory

Named datasets with full lineage — draft to promoted — built on record contracts and a quality bar every example must clear before import. You always know exactly what a model was trained on.

Curation gates that fail closed

Legal-safety review, privacy review, and tool-failure skips gate every candidate record, and eval holdouts are split from training data so the measurement stays honest.

Fine-tune on your infrastructure

Parameter-efficient (LoRA-class) supervised fine-tuning of local open-source models, run on the firm's own hardware. The dataset and the weights never leave your deployment.

Quality you can inspect

Analytics, run traces, and holdout evaluation show whether the model is actually getting better at your firm's work — not just different.

The persona holds

Training records enforce Lex's grounded, cite-your-sources, approval-gated behavior — so a firm-tuned model gets sharper at your practice without getting looser about the rules.

In your control

The dataset is your firm's curated work product; the weights are yours; training and serving run in your deployment. And a firm-tuned model is still just a model under EdgeLex governance — evidence gates and approval gates apply to it like any other.

See EdgeLex on your own terms.

We'll walk through self-hosting, model control, and your firm's workflows.