ArshPathan commited on
Commit
8da3892
·
1 Parent(s): 13bffd9

Upload folder using huggingface_hub

Browse files
training/MaterialForge_GRPO_Training.ipynb CHANGED
@@ -357,7 +357,7 @@
357
  "source": [
358
  "## 7. Load Model with Unsloth\n",
359
  "\n",
360
- "We use Unsloth for memory-efficient 4-bit QLoRA fine-tuning. This notebook now defaults to `Qwen/Qwen3-1.7B` for a more capable but still practical tool-using policy. If you have more VRAM, override the model name with an environment variable instead of editing the notebook repeatedly."
361
  ]
362
  },
363
  {
@@ -508,7 +508,7 @@
508
  "import torch\n",
509
  "from unsloth import FastLanguageModel\n",
510
  "\n",
511
- "MODEL_NAME = os.getenv(\"MATERIALFORGE_MODEL_NAME\", \"Qwen/Qwen3-1.7B\")\n",
512
  "MAX_SEQ_LENGTH = 4096\n",
513
  "\n",
514
  "model, tokenizer = FastLanguageModel.from_pretrained(\n",
 
357
  "source": [
358
  "## 7. Load Model with Unsloth\n",
359
  "\n",
360
+ "We use Unsloth for memory-efficient 4-bit QLoRA fine-tuning. This notebook now defaults to `Qwen/Qwen3-14B` for a stronger tool-using policy on larger hardware. If you want to try another checkpoint, override the model name with an environment variable instead of editing the notebook repeatedly."
361
  ]
362
  },
363
  {
 
508
  "import torch\n",
509
  "from unsloth import FastLanguageModel\n",
510
  "\n",
511
+ "MODEL_NAME = os.getenv(\"MATERIALFORGE_MODEL_NAME\", \"Qwen/Qwen3-14B\")\n",
512
  "MAX_SEQ_LENGTH = 4096\n",
513
  "\n",
514
  "model, tokenizer = FastLanguageModel.from_pretrained(\n",