forgelm 0.9.0

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

ForgeLM

A config-driven LLM fine-tuning toolkit for everyone — from solo researchers to enterprise platform teams. SFT → DPO → SimPO → KTO → ORPO → GRPO, with safety evaluation, EU AI Act compliance, and CI/CD-native design baked in. YAML in — fine-tuned model + audit artefacts out.

Use it interactively from a Jupyter notebook, drop it into a CI/CD pipeline, or run it from the terminal — the same YAML and the same Python API drive every entry point. Runs on Linux, macOS, and Windows.¹

---

Quick Start

pip install forgelm

# Fastest path: a bundled template that runs on a 12 GB GPU
forgelm quickstart customer-support

# Or generate a config interactively
forgelm --wizard

# Validate, fit-check, then train
forgelm --config my_config.yaml --dry-run
forgelm --config my_config.yaml --fit-check
forgelm --config my_config.yaml

# After training: chat, export, deploy
forgelm chat ./checkpoints/final_model
forgelm export ./checkpoints/final_model --quant q4_k_m
forgelm deploy ./checkpoints/final_model --target ollama

See the Quick Start Guide for the full walkthrough.

---

Why ForgeLM

• Config-driven. Behaviour is set in validated YAML — reproducible across notebooks, terminals, and CI runs with no hidden env-var flags.
• Full alignment stack. Every modern post-training method in one tool, one schema.
• Safety and compliance are first-class. Not an afterthought, not a separate product.
• CI/CD-native. Stable exit codes (0/1/2/3/4/5), JSON output, append-only audit log, deterministic dry-runs.
• Bring-your-own-data. PDF / DOCX / EPUB / Markdown → SFT-ready JSONL with a single command.
• Open source. Apache-2.0, no telemetry, no required cloud service.

---

Features

Training

• 6 trainer types: SFT, DPO, SimPO, KTO, ORPO, GRPO
• Memory-efficient methods: 4-bit QLoRA, DoRA, PiSSA, rsLoRA, GaLore
• Backends: Unsloth (2–5× faster) or standard Transformers
• Distributed: DeepSpeed ZeRO-2/3, FSDP, multi-GPU, MoE-aware (Qwen3, Mixtral, DeepSeek)
• Long-context: RoPE scaling, NEFTune, sliding-window attention, sample packing
• Multi-dataset mixing and synthetic data distillation (teacher → student)

Data Pipeline

• forgelm ingest — PDF / DOCX / EPUB / TXT / Markdown → SFT-ready JSONL, with token-aware and markdown-aware chunking
• forgelm audit — length, language, near-duplicate detection (SimHash + optional MinHash LSH), cross-split leakage, PII (TR / DE / FR / US-SSN, Luhn-validated), and a 9-family secrets scan
• PII masking on ingest (emails, phones, cards, IBAN, national IDs) and secrets masking before chunks land in the JSONL
• Croissant 1.0 dataset cards — the same JSON doubles as your EU AI Act Article 10 governance artefact

Evaluation & Safety

• Benchmarks via lm-evaluation-harness
• LLM-as-Judge scoring (OpenAI API or local model)
• Llama Guard safety classifier with S1–S14 harm categories, severity tiers, and cross-run trend tracking
• Auto-revert — runs that fail loss, benchmark, or safety thresholds are discarded before artefacts are written
• VRAM fit-check — pre-flight FITS / TIGHT / OOM / UNKNOWN estimator with concrete recommendations

Production & Deployment

• forgelm chat — streaming REPL with slash commands and optional safety routing
• forgelm export — GGUF export (6 quant levels) via llama-cpp-python
• forgelm deploy — generates Ollama, vLLM, TGI, or HF Endpoints configs
• Model merging (TIES, DARE, SLERP, linear) and auto-generated HF model cards
• Webhooks (Slack / Teams) and tracking via W&B / MLflow / TensorBoard
• Stable Python API: from forgelm import ForgeTrainer, audit_dataset, verify_audit_log, ... — every CLI surface has a typed entry point

---

Compliance & Safety

Most fine-tuning tools stop at "the model trained." ForgeLM produces the artefacts an auditor will ask for next:

• EU AI Act — auto-generated Annex IV technical documentation, Article 10 data governance, Article 14 human-oversight staging gate
• GDPR — forgelm purge (Article 17 right-to-erasure) and forgelm reverse-pii (Article 15 right-of-access)
• Model & log integrity — a SHA-256 manifest per trained model (forgelm verify-integrity) and a tamper-evident audit chain (forgelm verify-audit) give you a one-command proof-of-integrity before you ship
• Append-only audit log — HMAC-chained when FORGELM_AUDIT_SECRET is configured; every decision gate emits a structured event
• Supply-chain hardening — CycloneDX 1.5 SBOM per release, nightly pip-audit + bandit, gitleaks pre-commit
• ISO 27001 / SOC 2 alignment — software cannot be certified, but ForgeLM produces the change-management, data-lineage, and audit-trail evidence your deployer's auditor needs. See the Deployer Audit Guide.

Full details: Safety & Compliance Guide · Supply-Chain Security

---

Documentation

Topic	English	Türkçe
Quick Start	quickstart.md	—
Document Ingestion	ingestion.md	ingestion-tr.md
Dataset Audit	data_audit.md	data_audit-tr.md
Alignment (DPO / SimPO / KTO / GRPO)	alignment.md	—
Multi-Stage Pipelines	pipeline.md	pipeline-tr.md
CI/CD Integration	cicd_pipeline.md	cicd_pipeline-tr.md
Enterprise Deployment	enterprise_deployment.md	—
Safety & Compliance	safety_compliance.md	—
Troubleshooting & FAQ	troubleshooting.md	—
Architecture Reference	architecture.md	architecture-tr.md
Configuration Reference	configuration.md	configuration-tr.md
Product Strategy & Roadmap	product_strategy.md · roadmap.md	product_strategy-tr.md · roadmap-tr.md

---

Notebooks

Featured walkthroughs, runnable in Colab on a free T4 GPU:

• Quick Start — SFT Fine-Tuning
• GRPO Reasoning RL
• Safety Evaluation & Red-Teaming

See notebooks/ for the full set (DPO, KTO, multi-dataset, GaLore, synthetic data, post-training workflow, data curation).

---

Installation

# From PyPI
pip install forgelm

# From source
git clone https://github.com/HodeTech/ForgeLM.git
cd ForgeLM
pip install -e .

Prerequisites: Python 3.10+, torch>=2.4.0 (required by transformers>=5.3.0). Intel Macs (x86_64) are no longer supported — PyPI has no torch>=2.4 wheel for that platform. Platform-specific notes are in the installation guide.

Optional extras

pip install "forgelm[qlora]"            # 4-bit quantization (Linux)
pip install "forgelm[unsloth]"          # Unsloth backend (Linux)
pip install "forgelm[eval]"             # lm-evaluation-harness
pip install "forgelm[tracking]"         # W&B / MLflow
pip install "forgelm[distributed]"      # DeepSpeed
pip install "forgelm[merging]"          # model merging (TIES/DARE/SLERP — native, no extra deps)
pip install "forgelm[ingestion]"        # PDF / DOCX / EPUB / Markdown
pip install "forgelm[ingestion-scale]"  # MinHash LSH for large corpora
pip install "forgelm[ingestion-pii-ml]" # Presidio NER (also needs spaCy model)
pip install "forgelm[export]"           # GGUF via llama-cpp-python
pip install "forgelm[chat]"             # Rich terminal rendering

---

Docker

docker build -t forgelm --build-arg INSTALL_EVAL=true .

docker run --gpus all \
  -v $(pwd)/my_config.yaml:/workspace/config.yaml \
  -v $(pwd)/output:/workspace/output \
  forgelm --config /workspace/config.yaml

Multi-GPU and air-gapped deployment patterns are documented in the Enterprise Deployment Guide.

---

Contributing & License

Contributions are welcome — start with CONTRIBUTING.md and the engineering standards in docs/standards/.

Licensed under the Apache License 2.0.

1. qlora and unsloth extras depend on Linux-only upstream wheels; on macOS and Windows the install succeeds but those backends are skipped via a sys_platform == 'linux' marker. All other extras are cross-platform. ↩