Files
delentia-slm-jitna-v0.4/training_config/slm_jitna_v0.3.yaml
ModelHub XC 14fcab74b5 初始化项目,由ModelHub XC社区提供模型
Model: Delentia/delentia-slm-jitna-v0.4
Source: Original Platform
2026-08-06 15:55:17 +08:00

86 lines
2.7 KiB
YAML
Raw Permalink Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# 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)