Parameter	Value
Training Mode	SFT
Base Model	`Qwen/Qwen3.8-27B`
Learning Rate	0.0002
Epochs	2
Batch Size	1
Gradient Accumulation	16
Effective Batch Size	16
Max Sequence Length	2048
Optimizer	paged_adamw_8bit
LR Scheduler	cosine
Warmup Ratio	0.05
Weight Decay	0.01
Max Grad Norm	1.0
Seed	42
LoRA Rank (r)	32
LoRA Alpha	64
LoRA Dropout	0.05
Target Modules	up_proj, down_proj, gate_proj, k_proj, q_proj, v_proj, o_proj
GPU	NVIDIA GB10

Reproduce this training run

This model was trained with Merlina. Credentials are not included — Merlina will use your own HF_TOKEN and WANDB_API_KEY from .env or the form.

Paste this code into Merlina's Load Configuration → From Code to rebuild the exact training setup:

merlina-config-v1:H4sIAF_ghGoC_4WVS28jNwzHv0qgs8cZO27q9S0tUBRoctju9lQUAj3DmVGtV_SwNw32u5fUPOxFW_SmoSjqT-pHzruQBhO0kEAc3oUFg-IgfkajXXOqPl7QPqz31fb7H6pn9-uTWIkWYxOUT8pZcvwcQFll-7vG2U71OQBv3MUBArZ3XXDmDu5eMGhloUrsTGbjWtRripWgj-Lw-x8rEZsBDSkQZvS9T1PgagwsZhd5xhDL3ZvV7Hy1ie16u67F15U4QkRZLiIrp3F_kwtFczn5nOT_55spjnaBtKWQcSV4LYM4PGynNWg_0O7jbvpug_MUXRzqdf0d5xh6TCwla-RsRfbSB_cn19Jd7LzuIeG8Ps2L13lxnhduXFDJpogyORnhjGMdi4IE8STTm-fMfnz67dPTs3x-obMaIXBNJb0SFn11TVnYbCR61wykbsuVS80go_oLS437AK1CmyQ0TTZZlweWMaEn980j6YAvUqPt00DH691-tJBK49Oysam3uyIZ9awMcnLMwPTQ5bHIHLtE1iMyj_WaBGg40qFonEsD-RXdXC5jigeVmOGNmJjf6HJo8LqaLxty39PhDmhzJQKlK1XLdnp4sP39DMCnnz4TcebInDJzXiuKO4rk7xQQTFHRgY54Y5Hf1I0qWxOF0LaK6wWUQNEz4d65YKAIHlezzOROaClCoLvQwlFTApT06fZGrptsMWGTbjYZTrqwcTobKw14X8zvQlmSmJupX8dnWfgnSzO4iFaUs5YaiVDFGKEnsmaVI_gttpnK0RR2Ji3FhkQDE9W_LRfIKSo9L8Y0FYXejfCKb8SOmfmYHr1TWkskw9t_EEHyItKTHHbbkS_GUloSSMVmHi4QDDVWmT9z6-GZ6z6SWqDhXt4d-UUn_Wy4gG2Pc5I-x4E7asjHxYmmTD_OAHlEqglKqoODdnEYOvanlNT5tjhk5vESPUO3oMYv-8U7qnPf525x5g_5msGOLJRB8XEnf5Ev3OsnRF_8Zec3j1f4htx1xMjSAGMSF1T9kIiRBt5K5txFQfIA5ZERZtoaF4lzcdPj5NGcvFM2LVDxTCm9s_Q8wUPj3zCmcyRPvLSS_iLmIvdcYEYfjOQ23rCEDzeGbTF8mE0UVOky0LGq98VIbZpckAF52pyx3Hzlf9mPDVDyHgJVOVHP_NNjooJwSv-yWwo0DcOqXn9zd6OVp_YKGAen2wkySMlKZbxGQ9WCqaemQVZ-FKq_0UFxaCLQOPFKX7EoVJYRCuGqiiZrUrL3-baZpsiF0DL3qecnknhYvfLfar84hGzn_9nk8bCvqHmq8LCttugr6_hu8fVvMHUbuPQHAAA=

Prefer JSON? The same config, expanded

Save this to data/configs/<name>.json, or import it via the Load Configuration dialog.

json

{
  "_metadata": {
    "name": "Hemlock-Qwen3.8-27B-LoRA",
    "description": "Training configuration shared from a Merlina-trained model.",
    "tags": [],
    "schema": "merlina/training-config",
    "schema_version": 1,
    "merlina_version": "2.2.0"
  },
  "base_model": "Qwen/Qwen3.8-27B",
  "output_name": "Hemlock-Qwen3.8-27B-LoRA",
  "use_lora": true,
  "lora_r": 32,
  "lora_alpha": 64,
  "lora_dropout": 0.05,
  "target_modules": [
    "up_proj",
    "down_proj",
    "gate_proj",
    "k_proj",
    "q_proj",
    "v_proj",
    "o_proj"
  ],
  "modules_to_save": [],
  "lora_task_type": "CAUSAL_LM",
  "learning_rate": 0.0002,
  "num_epochs": 2,
  "batch_size": 1,
  "gradient_accumulation_steps": 16,
  "max_length": 2048,
  "max_prompt_length": 1024,
  "model_type": "auto",
  "training_mode": "sft",
  "beta": 0.1,
  "label_smoothing": 0.0,
  "gamma": 0.5,
  "dataset": {
    "source": {
      "source_type": "huggingface",
      "repo_id": "hemlang/Hemlock-SFT-combined",
      "split": "train",
      "streaming": false,
      "streaming_batch_size": 10000
    },
    "additional_sources": [],
    "format": {
      "format_type": "tokenizer",
      "enable_thinking": false,
      "auto_detect_thinking": true
    },
    "column_mapping": {
      "instruction": "prompt",
      "output": "chosen"
    },
    "convert_messages_format": true,
    "deduplicate": false,
    "dedupe_strategy": "prompt_chosen",
    "test_size": 0.02,
    "system_prompt_mode": "fill_empty",
    "training_mode": "sft"
  },
  "seed": 42,
  "max_grad_norm": 1.0,
  "warmup_ratio": 0.05,
  "eval_steps": 0.0,
  "use_4bit": false,
  "use_wandb": true,
  "push_to_hub": false,
  "merge_lora_before_upload": false,
  "hf_hub_private": false,
  "hf_namespace": "hemlang",
  "export_gguf": false,
  "gguf_quant_types": [
    "Q4_K_M"
  ],
  "keep_gguf_fp16": false,
  "shuffle_dataset": true,
  "weight_decay": 0.01,
  "lr_scheduler_type": "cosine",
  "gradient_checkpointing": true,
  "logging_steps": 1,
  "optimizer_type": "paged_adamw_8bit",
  "adam_beta1": 0.9,
  "adam_beta2": 0.999,
  "adam_epsilon": 1e-08,
  "adafactor_relative_step": false,
  "adafactor_scale_parameter": false,
  "adafactor_warmup_init": false,
  "adafactor_decay_rate": -0.8,
  "adafactor_clip_threshold": 1.0,
  "attn_implementation": "auto",
  "use_liger": false,
  "torch_compile": false,
  "eval_on_start": false,
  "multi_gpu_strategy": "auto",
  "wandb_project": "hemlock-qwen38",
  "wandb_run_name": "hemlock38-sft-r32-2ep-noeval"
}

Merlina on GitHub

hemlang/Hemlock-Qwen3.8-27B-LoRA

Hemlock-Qwen3.8-27B-LoRA

Training Configuration

Reproduce this training run

hemlang/
Hemlock-Qwen3.8-27B-LoRA