Solstice-1 / configs /sft3 /plan.json
wzebrowski's picture
Prepare Solstice-1 175M private release package
7bcd331 verified
Raw History Blame Contribute Delete
6.23 kB
{
"adaptation": {
"adapter_compute": "float32",
"alpha": 64,
"base_compute": "bfloat16",
"dropout": 0.0,
"expected_trainable_parameters": 3785856,
"layers": "all",
"method": "lora",
"rank": 32,
"targets": [
"attn.qkv_proj",
"attn.proj"
],
"trainable_token_ids": [
2,
3
]
},
"answer_max_tokens": {
"constraints": 320,
"dialogue": 256,
"general": 256,
"grounded": 160,
"knowledge": 80,
"rewrite": 320,
"skills": 160,
"social": 128
},
"authored_holdouts": {
"constraints": 128,
"dialogue": 128,
"skills": 192
},
"authored_quotas": {
"constraints": 4200,
"dialogue": 6800,
"skills": 12000
},
"baseline_cache": "${WORKSPACE}/pretraining/prepared/chat_adapter_search_20261108/baseline_cache",
"batches": [
8
],
"betas": [
0.9,
0.95
],
"checked_components": [
"knowledge",
"grounded",
"skills",
"constraints",
"dialogue"
],
"comparison_checkpoint": "${WORKSPACE}/reports/benchmarks/deep44-reliable-chat-167-fp32-20261106/choices/checkpoint",
"component_quotas": {
"constraints": 800,
"dialogue": 2000,
"general": 2400,
"grounded": 2000,
"skills": 800
},
"components": [
"general",
"knowledge",
"grounded",
"skills",
"constraints",
"dialogue",
"social",
"rewrite"
],
"epochs": 1,
"eps": 1e-08,
"eval_every": 64,
"eval_examples_per_component": 64,
"execution": "single-mac",
"final_lr": 2.0000000000000003e-06,
"format": "opensml-verified-chat-v2",
"fresh_mixture": {
"abstention": 0.1,
"dialogue": 0.15,
"grounded": 0.15,
"reasoning": 0.6
},
"fresh_weight": 0.4,
"generated_eval_per_component": 64,
"grad_accum": 8,
"grad_clip": 1.0,
"initial_checkpoint": "${WORKSPACE}/pretraining/runs/deep44_parent2000_generalize_05_replica_20261109/step_0000256_6074b98fe66b",
"keep_checkpoints": 6,
"kl_weight": 0.25,
"log_every": 10,
"max_cumulative_retention_loss_increase": 0.03,
"max_messages": 8,
"max_repetition_rate": 0.025,
"max_retention_loss_increase": 0.003,
"max_sequence_length": 2048,
"max_steps": 256,
"method": "Short checked reasoning",
"microbatch_rows": 4,
"min_domain_pass_cases": {},
"min_natural_stop": 0.9861111111111112,
"min_task_gain_cases": 4,
"minimum_source_tokens": 42000000000,
"original_pretraining_retention_loss": 2.8353193145245315,
"output": "${WORKSPACE}/pretraining/runs/deep44_chat_repair_04_20261109",
"peak_lr": 2e-05,
"prepared": "${WORKSPACE}/pretraining/runs/deep44_chat_repair_04_20261109/input/corpus",
"public_holdouts": {
"constraints_smol": 32,
"dialogue_smol": 64,
"general_smol": 64,
"grounded_no": 32,
"grounded_yes": 96,
"knowledge": 128,
"rewrite_smol": 32,
"social": 32
},
"public_quotas": {
"constraints_smol": 2000,
"dialogue_smol": 1200,
"general_smol": 4800,
"grounded_no": 2750,
"grounded_yes": 8250,
"knowledge": 4000,
"rewrite_smol": 1000,
"social": 1000
},
"qualitative_eval_per_component": 8,
"regression_patience": 2,
"repair_parent": "attempt5",
"replay_length": 512,
"replay_mixture": {
"constraints": 0.125,
"dialogue": 0.125,
"general": 0.1875,
"grounded": 0.125,
"knowledge": 0.1875,
"rewrite": 0.0625,
"skills": 0.125,
"social": 0.0625
},
"replay_objective_weight": 0.2,
"replay_policy": "prose-only",
"replay_rows_per_update": 16,
"retention": "${WORKSPACE}/pretraining/data/deep44_pretraining_validation_20261102",
"rows_per_group": 16,
"seed": 2026110915,
"selection_reference": "New 300-case development test, with historical 320-case and prose retention guards. Final 1000-case test is unopened until a single finalist is frozen.",
"sft_objective_weight": 0.8,
"sft_rows_per_update": 48,
"sft_stage": 2,
"source_checkpoint": "${WORKSPACE}/pretraining/runs/deep44_cooldown_2b_20261107/step_0320445_08b689895ed2",
"source_run": "${WORKSPACE}/pretraining/runs/deep44_cooldown_2b_20261107",
"sources": [
{
"file": "train.jsonl.gz",
"label": "helpsteer2",
"license": "cc-by-4.0",
"repo": "nvidia/HelpSteer2",
"revision": "990b2711a36180dd19d9c94b8627844866f8982a"
},
{
"file": "data/train-00000-of-00001-b42a775f407cee45.parquet",
"label": "oasst1",
"license": "apache-2.0",
"repo": "OpenAssistant/oasst1",
"revision": "fdf72ae0827c1cda404aff25b6603abec9e3399b"
},
{
"files": [
"data/smol-magpie-ultra/train-00000-of-00006.parquet",
"data/smol-magpie-ultra/train-00001-of-00006.parquet",
"data/smol-magpie-ultra/train-00002-of-00006.parquet",
"data/smol-magpie-ultra/train-00003-of-00006.parquet",
"data/smol-magpie-ultra/train-00004-of-00006.parquet",
"data/smol-magpie-ultra/train-00005-of-00006.parquet"
],
"label": "smol_magpie_ultra",
"license": "apache-2.0",
"repo": "HuggingFaceTB/smoltalk",
"revision": "5feaf2fd3ffca7c237fc38d1861bc30365d48ffa"
},
{
"label": "squad",
"license": "cc-by-sa-4.0",
"parquet_urls": [
"https://huggingface.co/datasets/rajpurkar/squad_v2/resolve/3ffb306f725f7d2ce8394bc1873b24868140c412/squad_v2/train-00000-of-00001.parquet"
],
"repo": "rajpurkar/squad_v2",
"revision": "3ffb306f725f7d2ce8394bc1873b24868140c412",
"split": "train"
},
{
"label": "squad_validation",
"license": "cc-by-sa-4.0",
"parquet_urls": [
"https://huggingface.co/datasets/rajpurkar/squad_v2/resolve/3ffb306f725f7d2ce8394bc1873b24868140c412/squad_v2/validation-00000-of-00001.parquet"
],
"repo": "rajpurkar/squad_v2",
"revision": "3ffb306f725f7d2ce8394bc1873b24868140c412",
"split": "validation"
}
],
"strategy": "Independent small adapter updates from parent2000 or preserved attempt5; verified text tasks and 60% replay objective with frozen-source probability preservation.",
"tokenizer": "${WORKSPACE}/reference/opensml-150m/tokenizer/bytebpe32k_v1",
"warmup_fraction": 0.1,
"weight_decay": 0.0
}