Kill_LIFE/docs/MISTRAL_FORGE_GUIDE.md

Mistral Forge Guide

How Forge reviews code in the Kill_LIFE / Mascarade ecosystem

Agent: Forge Agent ID: ag_019d1251023f73258b80ac73f90458f6 Model: codestral-latest (temperature 0.21) Domains: finetune, dataset, training, evaluation, benchmark, data

---

Overview

Forge is the code-oriented fine-tune and data agent of the Mascarade mesh. Its primary responsibilities are:

1. Code review using Codestral for PCB/embedded/SPICE domain code
2. Dataset validation and pipeline management for Mistral fine-tune jobs
3. Fine-tune job orchestration (upload, configure, launch, monitor)
4. Benchmark evaluation of base vs fine-tuned models

Forge operates at temperature 0.21 -- low enough for precise code generation, with enough margin for creative problem-solving in dataset augmentation.

---

Codestral Code Review Pipeline

FIM (Fill-in-the-Middle) completions

Codestral supports FIM completions for inline code suggestions, integrated into Mascarade via:

• Core route: /v1/api/providers/codestral/fim
• API facade: /api/providers/codestral/fim
• Endpoint: https://codestral.mistral.ai/v1/fim/completions

This was implemented in T-MS-023 (Lot 24, session 9) directly in the Mascarade active repo at /Users/electron/Documents/Projets/mascarade.

PCB review use case

Forge can review KiCad schematics, SPICE netlists, and embedded firmware through the dispatch system. It uses Codestral's code understanding to:

• Identify design rule violations in KiCad netlists
• Validate SPICE simulation parameters
• Review STM32/ESP32 firmware for common embedded pitfalls
• Check dataset quality for fine-tune pipelines

---

Fine-tune Pipeline

Dataset preparation tools

Tool	Location	Purpose
merge_datasets.sh	tools/mistral/merge_datasets.sh	Merge and deduplicate JSONL datasets
validate_dataset.py	tools/mistral/validate_dataset.py	Validate ChatML format, count examples
build_datasets.py	tools/mistral/build_datasets.py	Build domain-specific datasets
extract_hypnoled_datasets.py	tools/mistral/extract_hypnoled_datasets.py	Extract HypnoLED-specific training data

Dataset domains

Domain	Source files	Merged output	Status
KiCad	build_kicad_dataset.py outputs	datasets/kicad_merged.jsonl	Merged + validated
SPICE + Embedded	build_spice_dataset.py + build_embedded_dataset.py + build_stm32_dataset.py	datasets/spice_embedded_merged.jsonl	Merged + validated

Fine-tune pipeline flow

1. Build raw datasets
   build_datasets.py -> JSONL per domain

2. Merge and deduplicate
   merge_datasets.sh -> kicad_merged.jsonl, spice_embedded_merged.jsonl

3. Validate format
   validate_dataset.py -> ChatML format check, example count, dedup stats

4. Upload to Mistral
   mistral_studio_tui.sh --files-upload -> File IDs

5. Launch fine-tune job
   mistral_studio_tui.sh --finetune-create -> Job ID
   Hyperparameters: 100 steps, lr=1e-5

6. Monitor progress
   mistral_studio_tui.sh --finetune-list -> Status tracking

7. Validate fine-tuned model
   weekly_benchmark.sh -> Quality comparison vs baseline

Fine-tune targets

Model	Base	Target name	Domain	Status
ft:kicad-v1	open-mistral-7b	KiCad specialist	PCB, schematic, DRC	Pending (T-MS-010)
ft:spice-embedded-v1	codestral-latest	SPICE + Embedded specialist	Analog sim, firmware	Pending (T-MS-011)

---

Benchmark Pipeline

Prompt bank

Location: tools/evals/prompts/metier_100_benchmark.jsonl

100 domain-specific prompts:

• 20 KiCad prompts (schematic, PCB, DRC, BOM, scripting)
• 20 SPICE prompts (simulation, analysis, modeling)
• 20 Embedded prompts (STM32, ESP32, peripherals, RTOS)
• 20 IoT prompts (protocols, sensors, connectivity)
• 20 Mixed prompts (cross-domain integration)

Batch benchmark (T-MS-012)

Once fine-tuned models are available:

1. Run full benchmark on base model (codestral-latest, open-mistral-7b)
2. Run full benchmark on fine-tuned model (ft:kicad-v1, ft:spice-embedded-v1)
3. Compare quality scores per domain
4. Generate comparative report

# Base model benchmark
bash tools/evals/weekly_benchmark.sh --all --model codestral-latest

# Fine-tuned model benchmark (once available)
bash tools/evals/weekly_benchmark.sh --all --model ft:kicad-v1

# Compare
bash tools/evals/weekly_benchmark.sh --compare

---

Studio TUI Cockpit

Location: tools/cockpit/mistral_studio_tui.sh (referenced but created in Lot 24 T-MS-001)

The Studio TUI provides 14 actions for managing Mistral AI Studio resources:

• Agents management
• Files upload/list/delete
• Fine-tune create/list/monitor
• Batch jobs
• OCR (IA Documentaire)
• Audio (STT)
• Codestral (FIM + Chat)
• Logs

---

Dispatch via dispatch_to_agent.sh

Location: tools/ai/dispatch_to_agent.sh

Forge handles domains: finetune, dataset, training, evaluation, benchmark, data.

# Fine-tune pipeline task
bash tools/ai/dispatch_to_agent.sh --lot T-MS-010 --domain finetune

# Dataset validation
bash tools/ai/dispatch_to_agent.sh --lot T-MS-002 --domain dataset

# Benchmark evaluation
bash tools/ai/dispatch_to_agent.sh --lot T-MS-012 --domain benchmark

# Local mode (zero cost)
bash tools/ai/dispatch_to_agent.sh --lot T-MS-010 --domain finetune --local

---

Key files

File	Purpose
tools/mistral/merge_datasets.sh	Merge + deduplicate JSONL datasets
tools/mistral/validate_dataset.py	Validate ChatML format
tools/mistral/build_datasets.py	Build domain-specific datasets
tools/evals/weekly_benchmark.sh	Benchmark pipeline
tools/evals/prompts/metier_100_benchmark.jsonl	100 domain prompts
tools/ai/dispatch_to_agent.sh	Agent dispatch (Forge domains)
tools/mistral/beta_api_client.py	Mistral Beta API client
tools/mistral/mistral_client.py	Mistral API client