86 lines
2.7 KiB
YAML
86 lines
2.7 KiB
YAML
# Delentia SLM — JITNA v0.3 Cognitive OS Kernel Fine-tuning Configuration
|
||
# Base: Llama 3.1 8B (Apache 2.0, Thai-capable)
|
||
# Method: Unsloth QLoRA (4-bit) — optimized for T4 16GB / A100 40GB
|
||
# Format: TOON (Token-Oriented Object Notation) — ALGO-42
|
||
#
|
||
# Delta from v0.2:
|
||
# - Data mixing including: Delta Engine state deltas, Intent Loop correction flows, RCT 7 rules
|
||
# - Lowered learning rate (5.0e-5) to prevent Catastrophic Forgetting
|
||
# - Keep LoRA rank 32, alpha 64, with RSLoRA for format and logic convergence stability
|
||
|
||
model:
|
||
base_model: "unsloth/Meta-Llama-3.1-8B-bnb-4bit"
|
||
tokenizer: "unsloth/Meta-Llama-3.1-8B-bnb-4bit"
|
||
max_seq_length: 4096 # covers large intent chains and context history
|
||
dtype: null # auto-detect
|
||
load_in_4bit: true
|
||
|
||
lora:
|
||
r: 32
|
||
lora_alpha: 64
|
||
lora_dropout: 0
|
||
bias: "none"
|
||
use_rslora: true
|
||
target_modules:
|
||
- "q_proj"
|
||
- "k_proj"
|
||
- "v_proj"
|
||
- "o_proj"
|
||
- "gate_proj"
|
||
- "up_proj"
|
||
- "down_proj"
|
||
task_type: "CAUSAL_LM"
|
||
|
||
training:
|
||
dataset_path: "datasets/processed/jitna_pairs_v03.jsonl"
|
||
dataset_split: "train"
|
||
validation_split: 0.05
|
||
max_samples: null
|
||
|
||
per_device_train_batch_size: 1
|
||
gradient_accumulation_steps: 8
|
||
|
||
learning_rate: 5.0e-5 # Lowered from 1.0e-4 to accommodate mixed domain training smoothly
|
||
lr_scheduler_type: "cosine"
|
||
warmup_ratio: 0.05
|
||
num_train_epochs: 5
|
||
|
||
bf16: true
|
||
fp16: false
|
||
optim: "adamw_8bit"
|
||
weight_decay: 0.01
|
||
max_grad_norm: 0.3
|
||
|
||
output_dir: "models/checkpoints/v0.3_cognitive_kernel"
|
||
save_strategy: "epoch"
|
||
save_total_limit: 3
|
||
logging_steps: 10
|
||
|
||
evaluation_strategy: "epoch"
|
||
load_best_model_at_end: true
|
||
metric_for_best_model: "eval_loss"
|
||
|
||
chat_template: |
|
||
<|system|>
|
||
You are Delentia OS v0.3 — a constitutional AI operating under RCT v5 governance.
|
||
You process intents through the JITNA v3 protocol.
|
||
You respond in TOON format (Token-Oriented Object Notation) for token efficiency.
|
||
Your responses must be factual, safe, and PDPA-compliant.
|
||
Always provide FDIA scores when applicable (F = D^I × A).
|
||
For security-violating prompts, you must output a rejection state (FDIAScore: 0.00).
|
||
<|user|>
|
||
{{ user_intent }}
|
||
<|assistant|>
|
||
|
||
mlflow:
|
||
experiment_name: "delentia-slm-jitna-v0.3-cognitive"
|
||
tracking_uri: "https://delentia-delentia-agent-monitor.hf.space"
|
||
log_model: true
|
||
|
||
target_metrics:
|
||
jitna_compliance: 0.98 # >= 98% JITNA v3 schema compliance
|
||
toon_compliance: 0.95 # >= 95% TOON format compliance
|
||
fdia_avg: 0.895 # avg F score >= 0.895
|
||
hallucination_rate: 0.0028 # <= 0.28% factual errors (SignedAI consensus)
|
||
token_savings_pct: 10.0 # >= 10% token savings (adjusted for realistic v0.3 savings)
|