Step 1 · Measure before tuning
The rule-writer eval
40 plain-word strategies in English, Hinglish and Hindi, each with the rule an expert would write, plus five the engine cannot run yet, where the right answer is to say so. We score today's rule reader field by field: stock, indicator, lengths, stop, target, holding days and RSI levels.
Step 2 · Tune
How we post-train Artham
1
Measure
Run the current Artham and a frontier model on Raise's eval sets. The misses say what to train.
2
Collect
Traces from Dvar Evals, reviewer corrections, synthetic cases for rare intents and Hindi, frontier answers that pass the checks.
3
Supervised tuning
LoRA adapters on Artham per task, so one base serves several tasks.
4
Preference tuning
Pairs the checks and reviewers marked better or worse, so the model learns what to avoid, not just what to copy.
5
Gate
The same suite as the Evals tab. A version ships only if it is no worse on any critical check.
6
Release
A new Artham version under Raise's name, with a model card and an eval report Raise can publish.
| Task | Where it runs | Why it suits a tuned 7B | Ships as (placeholder) |
|---|---|---|---|
| Strategy rule writing | Backtester (workstream 1) | Short structured output a schema can check | Artham Rule-1 |
| Support intent and answers | Dhan Assist (workstream 2) | High volume, repetitive, Hinglish heavy | Artham Support-1 |
| Tool calling | fuzz, Dhan Assist | Picking the right data tool and arguments is learnable from traces | Artham Tools-1 |
| News drafting | ScanX | Fixed house style; facts checked against the source feed | Artham News-1 |
Step 3 · Gate
A tuned model ships only through the gate
The same 30 Dhan Assist chats as the Evals tab, graded again here. The answers are sample answers written to show typical behaviour; the grading is code and runs now.
Frontier model
Large hosted model, prompted
29/30
chats passing every check
Small model
7B open model, prompted
8/30
chats passing every check
Small model, tuned (Artham Support-1)
Same 7B after Dvar post-training
27/30
chats passing every check
Confirms before acting: 67% against 100% for the frontier model.
Protects personal data: 97% against 100% for the frontier model.
Cases passed: 7 points behind the frontier model.
So the next training round targets exactly these failures. Until it passes, Dhan Assist routes those answers to the frontier model (see it happen).
Step 4 · Release under Raise's name
A model card Raise can publish
Raise owns the weights trained on its data and the Artham name. Dvar does the work in the background and does not reuse Raise's data or weights for anyone else. Each release ships with this card and an eval report comparing it with the current Artham and a frontier model on Raise's own checks.
- • Where it trains: Raise's AWS account, where Artham already lives, or Dvar's environment.
- • Where it serves: Dvar inference in India, or Raise's own endpoint.
- • Cadence: a release about every 4 weeks once the first task is live.
- • Open points: Artham's base model licence, where its weights sit, GPU budget.
Model card · draft
Artham Rule-1
- Publisher
- Raise Financial Services
- Base
- Artham 7B
- Method
- LoRA SFT, then preference tuning
- Task
- Plain words → strategy rule (JSON)
- Languages
- English, Hinglish, Hindi
- Eval set
- Rule-writer v1, 40+ cases
- Today
- run the eval above
- Tuned
- measured at release
- Gate
- no worse on any critical check
Measured here: the rule-writer score of the built-in reader, and the grading of the sample chats. Not measured: no model was trained for this demo, so there is no tuned score yet; model names are placeholders. The first tuned release is planned for week 14 of the project plan.