configuration.md

Configuration Guide

ForgeLM uses YAML files for all configuration — declarative, version-controllable, and CI/CD-ready.

See config_template.yaml for a complete annotated example.

---

model

Field	Type	Default	Description
name_or_path	string	required	HuggingFace model ID or local path
max_length	int	2048	Maximum context length
load_in_4bit	bool	true	Enable QLoRA 4-bit NF4 quantization
backend	string	"transformers"	"transformers" or "unsloth" (2-5x faster, Linux only)
trust_remote_code	bool	false	Allow custom code from model repos. Security risk — only enable for models that require it
offline	bool	false	Air-gapped mode: no HF Hub calls. Models/datasets must be local
bnb_4bit_use_double_quant	bool	true	Double quantization for extra VRAM savings
bnb_4bit_quant_type	string	"nf4"	Quantization type ("nf4" or "fp4")
bnb_4bit_compute_dtype	string	"auto"	Compute dtype: "auto", "bfloat16", "float16", "float32"

model.moe (Optional — MoE models)

Field	Type	Default	Description
quantize_experts	bool	false	Quantize inactive expert weights to int8 for VRAM savings
experts_to_train	string	"all"	"all" or comma-separated expert indices (e.g., "0,1,2")

model.multimodal (Optional — VLM models)

Field	Type	Default	Description
enabled	bool	false	Enable vision-language model fine-tuning
image_column	string	"image"	Column name for image paths/URLs in dataset
text_column	string	"text"	Column name for text/captions

---

lora

Field	Type	Default	Description
r	int	8	LoRA rank. Higher = more parameters
alpha	int	16	LoRA scaling factor
dropout	float	0.1	Dropout probability
bias	string	"none"	"none", "all", or "lora_only"
method	string	"lora"	PEFT method: "lora", "dora", "pissa", "rslora"
use_dora	bool	false	Enable DoRA (Weight-Decomposed LoRA)
use_rslora	bool	false	Rank-stabilized LoRA (recommended for r>64)
target_modules	list	["q_proj", "v_proj"]	Model modules to apply LoRA
task_type	string	"CAUSAL_LM"	Task type for PEFT

---

training

Field	Type	Default	Description
output_dir	string	"./checkpoints"	Checkpoint save directory
final_model_dir	string	"final_model"	Subdirectory for final artifacts
merge_adapters	bool	false	Merge adapters into base model before saving
trainer_type	string	"sft"	"sft", "dpo", "simpo", "kto", "orpo", "grpo"
num_train_epochs	int	3	Number of training epochs
per_device_train_batch_size	int	4	Batch size per GPU
gradient_accumulation_steps	int	2	Steps to accumulate before backward pass
learning_rate	float	2e-5	Learning rate (lower for alignment: 5e-6)
warmup_ratio	float	0.1	Warmup proportion
weight_decay	float	0.01	AdamW weight decay
eval_steps	int	200	Evaluate every N steps
save_steps	int	200	Save checkpoint every N steps
save_total_limit	int	3	Max checkpoints to keep
packing	bool	false	Sequence packing (SFT only)
report_to	string	"tensorboard"	"tensorboard", "wandb", "mlflow", "none"
run_name	string	null	W&B/MLflow run name (auto-generated if null)

OOM Recovery

Automatically halves per_device_train_batch_size and doubles gradient_accumulation_steps on CUDA out-of-memory errors, preserving the effective batch size. Retries until the minimum batch size is reached.

Field	Type	Default	Description
oom_recovery	bool	false	Retry training with smaller batch size on CUDA OOM
oom_recovery_min_batch_size	int	1	Stop retrying when batch size reaches this value

Example:

training:
  per_device_train_batch_size: 8
  gradient_accumulation_steps: 2
  oom_recovery: true
  oom_recovery_min_batch_size: 1  # try down to batch_size=1 before failing

Effective batch size (per_device_train_batch_size × gradient_accumulation_steps) is preserved across retries. Each retry attempt is logged to the audit trail.

GaLore (Optimizer-Level Memory Optimization)

Field	Type	Default	Description
galore_enabled	bool	false	Enable GaLore gradient low-rank projection
galore_optim	string	"galore_adamw"	GaLore optimizer variant. One of: "galore_adamw", "galore_adamw_8bit", "galore_adafactor", "galore_adamw_layerwise", "galore_adamw_8bit_layerwise", "galore_adafactor_layerwise". _8bit halves optimizer-state VRAM; _layerwise cuts peak VRAM by recomputing per-layer.
galore_rank	int	128	Rank for gradient projection
galore_update_proj_gap	int	200	Steps between projection updates
galore_scale	float	0.25	GaLore scaling factor
galore_proj_type	string	"std"	Projection type: "std", "reverse_std", "right", "left", "full"
galore_target_modules	Optional[List[str]]	null	Module-name regex patterns GaLore is applied to. null falls back to [r".*.attn.*", r".*.mlp.*"] (attention + MLP layers).

Long-Context Training

Field	Type	Default	Description
rope_scaling	Optional[Dict[str, Any]]	null	RoPE scaling method dict ({"type": "linear", "factor": 2.0} etc.). Supported types: "linear", "dynamic", "yarn", "longrope".
neftune_noise_alpha	float	null	NEFTune noise injection alpha (e.g., 5.0)
sliding_window_attention	int	null	Sliding window attention size in tokens
sample_packing	bool	false	Pack multiple short samples into full-length sequences

GPU Cost Estimation

Field	Type	Default	Description
gpu_cost_per_hour	float	null	Custom GPU cost rate (USD/hour). Auto-detected from GPU model if null

Alignment Parameters

Field	Type	Default	Used By
dpo_beta	float	0.1	DPO temperature
simpo_gamma	float	0.5	SimPO margin term
simpo_beta	float	2.0	SimPO scaling
kto_beta	float	0.1	KTO loss parameter
orpo_beta	float	0.1	ORPO odds ratio weight
grpo_num_generations	int	4	GRPO: responses per prompt
grpo_max_completion_length	int	512	GRPO: max tokens per completion (legacy alias grpo_max_new_tokens accepted)
grpo_reward_model	string	null	GRPO: reward model path (HF or local)

---

data

Field	Type	Default	Description
dataset_name_or_path	string	required	HF dataset ID or local JSONL path
extra_datasets	list	null	Additional datasets to mix in
mix_ratio	list	null	Weight per dataset (e.g., [0.7, 0.3])
shuffle	bool	true	Shuffle training data
clean_text	bool	true	Strip extra whitespace
add_eos	bool	true	Add EOS token to sequences

data.governance (Optional — EU AI Act Art. 10)

Field	Type	Default	Description
collection_method	string	""	How data was collected
annotation_process	string	""	Annotation methodology
known_biases	string	""	Known dataset biases
personal_data_included	bool	false	Contains personal data
dpia_completed	bool	false	Data Protection Impact Assessment done

---

evaluation (Optional)

Field	Type	Default	Description
auto_revert	bool	false	Delete model if evaluation fails
max_acceptable_loss	float	null	Hard ceiling for eval_loss
baseline_loss	float	null	Computed automatically if null
require_human_approval	bool	false	Pause for human review (exit code 4)

evaluation.benchmark (Optional)

Field	Type	Default	Description
enabled	bool	false	Enable lm-eval-harness benchmarks
tasks	list	[]	Task names (e.g., ["arc_easy", "hellaswag"])
num_fewshot	int	null	Few-shot examples (task default)
batch_size	string	"auto"	Evaluation batch size
limit	int	null	Samples per task (for quick checks)
min_score	float	null	Minimum average accuracy

evaluation.safety (Optional)

Field	Type	Default	Description
enabled	bool	false	Enable safety classifier evaluation
classifier	string	"meta-llama/Llama-Guard-3-8B"	Safety classifier model
test_prompts	string	"safety_prompts.jsonl"	Adversarial test prompts file. Built-in sets in configs/safety_prompts/
max_safety_regression	float	0.05	Max allowed unsafe ratio (binary gate)
scoring	string	"binary"	Scoring mode: "binary" or "confidence_weighted"
min_safety_score	float	null	Weighted score threshold (0.0-1.0). Used when scoring="confidence_weighted"
min_classifier_confidence	float	0.7	Flag responses below this confidence for manual review
track_categories	bool	false	Parse Llama Guard S1-S14 harm categories
severity_thresholds	dict	null	Per-severity limits: {"critical": 0, "high": 0.01, "medium": 0.05}
batch_size	int	8	Batched generation size for safety evaluation. 1 disables batching; raise for throughput on large VRAM, lower to reduce OOM risk on small VRAM.

evaluation.llm_judge (Optional)

Field	Type	Default	Description
enabled	bool	false	Enable LLM-as-Judge scoring
judge_model	string	"gpt-4o"	Judge model (API or local path)
judge_api_key_env	string	null	Env var name for API key (null = local)
judge_api_base	string	null	Override the judge API base URL (Azure OpenAI, self-hosted vLLM, OpenAI-compatible gateway, e.g. https://api.together.xyz/v1). When unset, the SDK default endpoint is used.
eval_dataset	string	"eval_prompts.jsonl"	Evaluation prompts file
min_score	float	5.0	Minimum average score (1-10)
batch_size	int	8	Number of (prompt, completion) pairs scored per LLM-judge round. 1 disables batching.

Deprecated: evaluation.staging_ttl_days is superseded by retention.staging_ttl_days. The legacy key is alias-forwarded with a DeprecationWarning during the v0.5.5 → v0.6.x window and removed in v0.7.0. See release.md.

---

retention (Optional — GDPR Article 17 erasure horizons)

Defines maximum retention horizons for compliance, training, and evaluation artefacts. Horizons honour GDPR Article 5(1)(e) "storage limitation" and Article 17 "right to erasure" deadlines. The enforce knob switches between log-only, warning, and hard-block modes so a regulated CI gate cannot silently extend the retention horizon by re-using a stale workspace.

Field	Type	Default	Description
audit_log_retention_days	int	1825 (~5 years)	Days to retain audit_log.jsonl before flagging it as overdue under Article 5(1)(e). Set to 0 to retain indefinitely (Article 17(3)(b) defence).
staging_ttl_days	int	7	Days to retain final_model.staging.<run_id>/ after a forgelm reject decision before scheduled cleanup. Set to 0 to retain indefinitely. Replaces the deprecated evaluation.staging_ttl_days; both keys accepted with identical values during the v0.5.5 → v0.6.x deprecation window.
ephemeral_artefact_retention_days	int	90	Days to retain compliance bundles, data audit reports, and other run-scoped derived artefacts. Set to 0 to retain indefinitely.
raw_documents_retention_days	int	90	Days to retain ingested raw documents (PDF / DOCX / EPUB / TXT / Markdown) under the operator's ingestion-output directory. Set to 0 to retain indefinitely.
enforce	string	"log_only"	Policy enforcement mode: "log_only" (audit-log only), "warn_on_excess" (structured stderr warning), "block_on_excess" (abort trainer pre-flight with EXIT_EVAL_FAILURE = 3).

Deprecation: evaluation.staging_ttl_days is deprecated as of v0.5.5 in favour of retention.staging_ttl_days. The legacy key is alias-forwarded with a DeprecationWarning until v0.7.0. See release.md for the full deprecation cadence policy.

---

webhook (Optional)

Field	Type	Default	Description
url	string	null	Webhook destination URL
url_env	string	null	Env var name containing URL
notify_on_start	bool	true	Notify on training start
notify_on_success	bool	true	Notify on success
notify_on_failure	bool	true	Notify on failure
timeout	int	10	HTTP request timeout (seconds). Clamped to ≥ 1s by the notifier. Default raised to 10s in v0.5.5 (was 5s) — Slack/Teams gateway latency spikes regularly cross 5s in production, and a webhook timeout silently degrades the audit chain (webhook failure is best-effort).
allow_private_destinations	bool	false	Opt in to webhooks pointing at RFC1918 / loopback / link-local hosts (in-cluster Slack proxy, on-prem Teams gateway). Defaults to public-internet only — SSRF guard
tls_ca_bundle	string	null	Path to a custom CA bundle forwarded to requests as verify= (e.g. corporate MITM CA). When unset, certifi's bundled store is used

---

distributed (Optional)

Field	Type	Default	Description
strategy	string	null	"deepspeed" or "fsdp" (null = single GPU)
deepspeed_config	string	null	Preset ("zero2", "zero3", "zero3_offload") or JSON path
fsdp_strategy	string	"full_shard"	"full_shard", "shard_grad_op", "hybrid_shard", "no_shard"
fsdp_auto_wrap	bool	true	Auto-wrap transformer layers
fsdp_offload	bool	false	Offload parameters to CPU
fsdp_backward_prefetch	string	"backward_pre"	"backward_pre" or "backward_post"
fsdp_state_dict_type	string	"FULL_STATE_DICT"	"FULL_STATE_DICT" or "SHARDED_STATE_DICT"

---

merge (Optional)

Field	Type	Default	Description
enabled	bool	false	Enable model merging
method	string	"ties"	"ties", "dare", "slerp", "linear"
models	list	[]	List of {path, weight} dicts
output_dir	string	"./merged_model"	Output directory

---

compliance (Optional — EU AI Act Art. 11 + Annex IV)

Field	Type	Default	Description
provider_name	string	""	Organization name
provider_contact	string	""	Contact email
system_name	string	""	AI system name
intended_purpose	string	""	What the model is for
known_limitations	string	""	What it should not be used for
system_version	string	""	Version identifier
risk_classification	string	"minimal-risk"	One of the 5 EU AI Act RiskTier values: "unknown" (pre-classification placeholder), "minimal-risk", "limited-risk", "high-risk" (Article 6 — full Annex IV documentation), "unacceptable" (Article 5 prohibited practice — emits a startup banner).

---

risk_assessment (Optional — EU AI Act Art. 9)

Field	Type	Default	Description
intended_use	string	""	Intended use description
foreseeable_misuse	list	[]	List of misuse scenarios
risk_category	string	"minimal-risk"	Same 5 RiskTier values as compliance.risk_classification: "unknown", "minimal-risk", "limited-risk", "high-risk", "unacceptable". Drives auto-revert thresholds and Annex IV gating.
mitigation_measures	list	[]	Risk mitigation measures
vulnerable_groups_considered	bool	false	Impact on vulnerable groups assessed

---

monitoring (Optional — EU AI Act Art. 12+17)

Field	Type	Default	Description
enabled	bool	false	Enable monitoring hooks
endpoint	string	""	Monitoring webhook URL
endpoint_env	string	null	Env var name for endpoint
metrics_export	string	"none"	"none", "prometheus", "datadog", "custom_webhook"
alert_on_drift	bool	true	Alert on model drift
check_interval_hours	int	24	Monitoring check interval

---

synthetic (Optional — Synthetic Data Generation)

Field	Type	Default	Description
enabled	bool	false	Enable teacher → student synthetic-data generation.
teacher_model	string	""	HF Hub ID or API model name (e.g. gpt-4o, meta-llama/Llama-3-70B).
teacher_backend	string	"api"	One of "api" (OpenAI/Anthropic-compatible), "local" (HF in-process), "file" (read pre-generated JSONL).
api_base	string	""	API endpoint, e.g. https://api.openai.com/v1 or self-hosted vLLM gateway.
api_key	Optional[str]	null	Inline API key. Prefer api_key_env to avoid committing secrets — when set inline, the value is ***REDACTED*** in serialized config.
api_key_env	Optional[str]	null	Env var name carrying the API key (e.g. OPENAI_API_KEY).
api_delay	float	0.5	Seconds between teacher calls (rate limiting).
api_timeout	int	60	Per-call API timeout in seconds.
seed_file	string	""	Path to seed prompts file (JSONL or plain text, one prompt per line).
seed_prompts	List[str]	[]	Inline seed prompts (alternative to seed_file).
system_prompt	string	""	System prompt prepended on every teacher call.
max_new_tokens	int	1024	Max tokens per teacher response.
temperature	float	0.7	Sampling temperature passed to the teacher.
output_file	string	"synthetic_data.jsonl"	Output JSONL file path.
output_format	string	"messages"	One of "messages" (chat-style array), "instruction" (Alpaca-style), "chatml", "prompt_response".

---

auth (Optional)

Field	Type	Default	Description
hf_token	string	null	HuggingFace token (prefer HUGGINGFACE_TOKEN env var)

---

pipeline (Optional — Multi-Stage Training Chains, Phase 14)

Chains 2+ training stages (typically SFT → DPO → GRPO) into one config-driven run with auto-chaining, per-stage gates, crash-safe resume, and a chain-level Annex IV manifest. When omitted, ForgeLM behaves byte-identically to a v0.6.0 single-stage run; the orchestrator module is not imported. Full operator walkthrough: Multi-Stage Pipelines guide.

Field	Type	Default	Description
output_dir	string	"./pipeline_run"	Root directory for chain-level artefacts: pipeline_state.json, compliance/pipeline_manifest.json, and the pipeline-scoped audit_log.jsonl. Per-stage trainer artefacts continue to live under each stage's own training.output_dir.
stages	List[PipelineStage]	[] (required: ≥ 1)	Ordered list of stages. Each stage's model.name_or_path is auto-set to the previous stage's training.output_dir/final_model unless the stage supplies an explicit model: block.

pipeline.stages[].* — PipelineStage fields

A PipelineStage is a per-stage override layered onto the root config. Section-wholesale inheritance: omitting a block inherits root's wholesale; supplying a block REPLACES root's wholesale (no deep-merge).

Field	Type	Default	Description
name	string	— (required)	Stage identifier matching ^[a-z0-9_]{1,32}$. Unique within the pipeline. Used as the identifier in --stage <name>, --resume-from <name>, audit-log payloads, and per-stage manifest entries.
model	Optional[ModelConfig]	null	Per-stage override of the root model: block. When null, auto-chains from the previous stage's final_model (or root for stage 0). When set, disables the auto-chain for that stage (operator escape hatch).
lora	Optional[LoraConfig]	null	Per-stage LoRA config. Inherits root wholesale when null.
training	Optional[TrainingConfig]	null	Per-stage training config. Inherits root wholesale when null. When supplied, trainer_type MUST be set explicitly — every stage records its alignment paradigm in the manifest for audit clarity.
data	Optional[DataConfig]	null	Per-stage data config. Inherits root wholesale when null; per-stage override is the norm because each stage typically consumes a different dataset (SFT/DPO/preference/etc.).
evaluation	Optional[EvaluationConfig]	null	Per-stage gates (loss thresholds, auto_revert, safety, judge, human-approval). Each stage may independently configure its gate.

Root-only sections — rejected at the stage level with EXIT_CONFIG_ERROR (1): distributed, webhook, compliance, risk_assessment, monitoring, retention, synthetic, merge, auth. These are pipeline-level concerns (distributed strategy stays consistent across the run; compliance metadata covers the whole chain; etc.).

Example

# Root defaults — inherited by stages that omit a block.
model: { name_or_path: "meta-llama/Llama-3-8B" }
lora: { r: 8, alpha: 16 }
training: { trainer_type: "sft", output_dir: "./placeholder" }
data: { dataset_name_or_path: "./placeholder.jsonl" }

pipeline:
  output_dir: "./pipeline_run"
  stages:
    - name: sft_stage
      training: { trainer_type: "sft", output_dir: "./pipeline_run/stage1_sft" }
      data: { dataset_name_or_path: "./data/sft.jsonl" }
    - name: dpo_stage
      training: { trainer_type: "dpo", output_dir: "./pipeline_run/stage2_dpo", dpo_beta: 0.1 }
      data: { dataset_name_or_path: "./data/preferences.jsonl" }
    - name: grpo_stage
      training: { trainer_type: "grpo", output_dir: "./pipeline_run/stage3_grpo" }
      data: { dataset_name_or_path: "./data/math_prompts.jsonl" }

CLI surface

Flag	Effect
--stage <name>	Run only the named stage in isolation (audit / re-run scenarios). Auto-chains from the previous stage's on-disk output.
--resume-from <name>	Resume from the named stage onward; already-completed (or human-approved gated) stages with on-disk output are skipped.
--force-resume	Accept a pipeline_config_hash mismatch on resume (logged + audited via pipeline.force_resume). Stage topology mismatch (count / names / order) is refused even with this flag.
--input-model <path>	Operator escape hatch — overrides the auto-chained model for the --stage target. Audit-logged with input_source: cli_override.
--dry-run	Validates every stage's merged config + cross-stage chain integrity + training.output_dir collision check before any GPU is allocated; collects all errors before exiting.

The --fit-check, --merge, --generate-data, --compliance-export, --benchmark-only flags are single-stage operations and are rejected at dispatch time when a pipeline: block is present — drop the pipeline: block or remove the flag.

Verifier

forgelm verify-annex-iv --pipeline <pipeline.output_dir>

Validates the chain-level manifest's structural fields, chain-integrity (every stage with input_source: chain matches its immediate predecessor's output_model), per-stage training_manifest.json existence, and stopped_at / running-status consistency. Exit 0 on clean manifest, 1 on config / chain violation, 2 on runtime I/O failure.