Files
delentia-slm-jitna-v0.4/training_config/slm_jitna_executor.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

81 lines
2.4 KiB
YAML
Raw 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 — The Executor (slm-jitna-agentic) LoRA Configuration
# Purpose: Function Calling / Structured JSON Output
# Priority: #1 in 4-Pillar Architecture (unlocks Intent Memory Loop)
#
# The Executor MUST produce pure JSON with zero natural language contamination.
# It converts user intents into machine-executable tool-call payloads.
model:
base_model: "Delentia/delentia-slm-jitna-v0.4"
tokenizer: "Delentia/delentia-slm-jitna-v0.4"
max_seq_length: 4096
dtype: null
load_in_4bit: true
lora:
r: 32
lora_alpha: 64 # α = 2r for optimal learning rate control
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_executor_pairs.parquet"
dataset_split: "train"
validation_split: 0.05
max_samples: null
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 3.0e-5 # Lower LR to force precise JSON structure learning
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/executor_agentic"
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 The Executor (slm-jitna-agentic) — a specialized LoRA adapter within the Delentia OS 1+4 Pillar Architecture. Your ONLY purpose is to convert user intents into machine-executable JSON payloads. You must NEVER produce natural language explanations. Output ONLY valid JSON — no markdown, no text, no comments. Your output must pass json.loads() without error.
<|user|>
{{ user_intent }}
<|assistant|>
pillar_type: "executor"
adapter_name: "jitna_executor_v1"
adapter_save_path: "models/adapters/jitna_executor_v1"
mlflow:
experiment_name: "delentia-slm-executor-agentic"
tracking_uri: "https://delentia-delentia-agent-monitor.hf.space"
log_model: true
target_metrics:
json_validity: 0.99 # >= 99% valid JSON output
tool_call_accuracy: 0.95 # >= 95% correct tool name + args
fdia_avg: 0.90 # avg F score >= 0.90
hallucination_rate: 0.02 # <= 2% factual errors