Text Generation
PEFT
Safetensors
English
qwen2
marine-science
oceanography
aquaculture
coastal-restoration
environmental-compliance
dfo-statutory
qwen
deepseek-r1-distill
lora
vllm
awq
cuda
aboveboard-benchmark
conversational
Eval Results (legacy)
Instructions to use Coralfil/Atlantis-Pyramid-v2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use Coralfil/Atlantis-Pyramid-v2 with PEFT:
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-32B-Instruct") model = PeftModel.from_pretrained(base_model, "Coralfil/Atlantis-Pyramid-v2") - Notebooks
- Google Colab
- Kaggle
Download training_args.json from Coralfil/Atlantis-Pyramid-v2: direct link, hf CLI and curl.
- Browser
- Download file 6.97 kB
-
https://huggingface.co/Coralfil/Atlantis-Pyramid-v2/resolve/main/training_args.json
- Command line
-
hf download hf://Coralfil/Atlantis-Pyramid-v2/training_args.json
-
curl -L -o training_args.json https://huggingface.co/Coralfil/Atlantis-Pyramid-v2/resolve/main/training_args.json
6.97 kB
| { | |
| "model_name": "Atlantis-Pyramid v2", | |
| "base_model": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B", | |
| "budget": { | |
| "total_authorized_spend_usd": 500.0, | |
| "allocation": { | |
| "a100_training_allocation_usd": 175.0, | |
| "a100_hourly_rate_usd": 1.29, | |
| "a100_estimated_compute_hours": 135.6, | |
| "a10_serving_allocation_usd": 325.0, | |
| "a10_hourly_rate_usd": 0.60, | |
| "a10_estimated_serving_hours": 541.6 | |
| }, | |
| "cost_optimization_policy": "Enforce 20-minute idle watchdog shutdown on A10 inference instance. Prevent 24/7 idle billing." | |
| }, | |
| "sequential_training": { | |
| "enabled": true, | |
| "strategy": "sequential_lora_adapters", | |
| "rationale": "Train distinct LoRA adapters sequentially across DeepSeek-R1 reasoning, Qwen2.5 direct instruction, and Mistral European outlier architectures to gauge cross-foundation synergy without weight corruption.", | |
| "models": [ | |
| { | |
| "id": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B", | |
| "license": "MIT", | |
| "context_window": 131072, | |
| "role": "Primary scientific deep reasoning with native chain of thought tags", | |
| "output_dir": "D:/Projects/os/atlantis/builds/atlantis-pyramid-32b/outputs/pyramid-32b-deepseek-r1-lora", | |
| "s3_prefix": "models/atlantis-pyramid-v2/lora-deepseek-r1-32b" | |
| }, | |
| { | |
| "id": "Qwen/Qwen2.5-32B-Instruct", | |
| "license": "Apache 2.0", | |
| "context_window": 131072, | |
| "role": "Direct instruction and high capacity operational co foundation", | |
| "output_dir": "D:/Projects/os/atlantis/builds/atlantis-pyramid-32b/outputs/pyramid-32b-qwen25-lora", | |
| "s3_prefix": "models/atlantis-pyramid-v2/lora-qwen25-32b" | |
| }, | |
| { | |
| "id": "mistralai/Mistral-Small-Instruct-2409", | |
| "license": "Apache 2.0", | |
| "context_window": 32768, | |
| "role": "European sovereign model outlier for cross architecture synergy validation", | |
| "output_dir": "D:/Projects/os/atlantis/builds/atlantis-pyramid-32b/outputs/pyramid-24b-mistral-lora", | |
| "s3_prefix": "models/atlantis-pyramid-v2/lora-mistral-small-24b" | |
| } | |
| ] | |
| }, | |
| "lora": { | |
| "r": 16, | |
| "lora_alpha": 32, | |
| "lora_dropout": 0.05, | |
| "bias": "none", | |
| "target_modules": [ | |
| "q_proj", | |
| "k_proj", | |
| "v_proj", | |
| "o_proj" | |
| ], | |
| "task_type": "CAUSAL_LM" | |
| }, | |
| "training": { | |
| "per_device_train_batch_size": 1, | |
| "gradient_accumulation_steps": 16, | |
| "learning_rate": 0.00015, | |
| "lr_scheduler_type": "cosine", | |
| "warmup_ratio": 0.05, | |
| "num_train_epochs": 3, | |
| "max_steps": 500, | |
| "weight_decay": 0.01, | |
| "optim": "paged_adamw_8bit", | |
| "fp16": false, | |
| "bf16": true, | |
| "gradient_checkpointing": true, | |
| "logging_steps": 5, | |
| "save_strategy": "steps", | |
| "save_steps": 50, | |
| "save_total_limit": 3 | |
| }, | |
| "dataset": { | |
| "train_file": "D:/Projects/os/atlantis/builds/atlantis-pyramid-32b/dataset/atlantis_pyramid_sft_packed.jsonl", | |
| "curated_file": "D:/Projects/os/atlantis/builds/atlantis-pyramid-32b/dataset/atlantis_pyramid_sft_curated.jsonl", | |
| "dpo_file": "D:/Projects/os/atlantis/builds/atlantis-pyramid-32b/dataset/atlantis_pyramid_dpo.jsonl", | |
| "max_seq_length": 131072, | |
| "packing": true | |
| }, | |
| "cache_strategy": { | |
| "automatic_prefix_caching": true, | |
| "chunked_prefill": true, | |
| "fp8_kv_cache": true, | |
| "torch_inductor_cache": "/workspace/.cache/torch_inductor", | |
| "vllm_cache_root": "/workspace/.cache/vllm", | |
| "lazy_module_loading": true | |
| }, | |
| "a10_serving_profile": { | |
| "instance_type": "gpu_1x_a10", | |
| "vram_gb": 24, | |
| "quantization": "awq_int4", | |
| "kv_cache_dtype": "fp8", | |
| "max_model_len": 16384, | |
| "gpu_memory_utilization": 0.85, | |
| "idle_watchdog_timeout_minutes": 20 | |
| }, | |
| "evaluation": { | |
| "suite": "AboveBoard 52-Scenario Marine Suite", | |
| "benchmark_file": "D:/Projects/os/data/corpora/canadian_marine/aboveboard_benchmark.json", | |
| "scoring_standard": "CRF-FS-ABOVEBOARD-52", | |
| "categories": [ | |
| "Category A: Statutory Marine Compliance (20 scenarios)", | |
| "Category B: Physical Oceanography and Biogeochemistry (18 scenarios)", | |
| "Category C: Benthic Restoration and Structural Biomaterials (14 scenarios)" | |
| ], | |
| "pass_threshold": 0.88, | |
| "safety_guardrails": { | |
| "prebiotic_chemistry_only": true, | |
| "prohibit_antibiotics": true, | |
| "prohibit_probiotics": true, | |
| "fisheries_act_section_36_enforced": true, | |
| "deter_biological_pathogen_engineering": true, | |
| "deter_toxic_dumping_and_leaching": true, | |
| "prohibit_unauthorized_mpa_core_disturbance": true | |
| } | |
| }, | |
| "frontier_safety_framework": { | |
| "version": "CFSF-v1.0-Xray", | |
| "policy_document": "D:/Projects/os/atlantis/builds/atlantis-pyramid-32b/coralfil_frontier_safety_framework_v1.0_Xray_2026-10-05T11-30-00.md", | |
| "security_level_mandate": "CSL-2+", | |
| "capability_thresholds": { | |
| "domain_a_ecotoxicity_cbrn": { | |
| "tracked_capability_level": "M-TCL-A1", | |
| "critical_capability_level": "M-CCL-A1", | |
| "alert_threshold_ratio": 0.70, | |
| "mitigation": "DPO refusal steering, strict prohibition of open-water antibiotics, probiotics, and toxic mineral leaching" | |
| }, | |
| "domain_b_maritime_cyber": { | |
| "tracked_capability_level": "M-TCL-B1", | |
| "critical_capability_level": "M-CCL-B1", | |
| "alert_threshold_ratio": 0.70, | |
| "mitigation": "Read-only sandbox, denial of socket creation and port scanning, telemetry signature validation" | |
| }, | |
| "domain_c_agentic_stealth_misalignment": { | |
| "tracked_capability_level": "M-TCL-C1", | |
| "critical_capability_level": "M-CCL-C1", | |
| "alert_threshold_ratio": 0.70, | |
| "mitigation": "Arize Phoenix chain-of-thought inspection for deceptive alignment, tool affordance clamping" | |
| }, | |
| "domain_d_ml_rd_exfiltration": { | |
| "tracked_capability_level": "M-TCL-D1", | |
| "critical_capability_level": "M-CCL-D1", | |
| "alert_threshold_ratio": 0.70, | |
| "mitigation": "Non-destructive S3 versioning, model weight HMAC signing, 20-minute idle watchdog shutdown" | |
| } | |
| }, | |
| "safety_case_required_at_ccl": true | |
| }, | |
| "observability": { | |
| "arize_phoenix": { | |
| "enabled": true, | |
| "endpoint": "http://localhost:6006", | |
| "project_name": "atlantis-pyramid-v2" | |
| }, | |
| "wandb": { | |
| "enabled": true, | |
| "project": "atlantis-pyramid-v2", | |
| "entity": "coralfil" | |
| } | |
| }, | |
| "storage_sync": { | |
| "target_s3_bucket": "f46ad631-a75f-4cd1-ae30-55973c5ecb55", | |
| "target_s3_prefix": "models/atlantis-pyramid-v2/lora", | |
| "region": "ca-central-1" | |
| }, | |
| "export_recipes": { | |
| "awq_int4": { | |
| "quant_method": "awq", | |
| "bits": 4, | |
| "group_size": 128, | |
| "zero_point": true, | |
| "version": "GEMM", | |
| "calib_data": "dataset/atlantis_pyramid_sft_curated.jsonl" | |
| }, | |
| "fp8": { | |
| "quant_method": "fp8", | |
| "format": "e4m3fn", | |
| "kv_cache_dtype": "fp8_e4m3", | |
| "activation_scheme": "dynamic" | |
| } | |
| } | |
| } |