Upload folder using huggingface_hub

.gitattributes

@@ -1,35 +1 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
 *.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,30 @@
----
-license: mit
----
+# PlutoLite-v1
+
+PlutoLite-v1 是基于 [google/gemma-4-E2B](https://huggingface.co/google/gemma-4-E2B) 的 Python 代码生成 / 函数补全 QLoRA 适配器。
+
+发布内容为 PEFT LoRA 权重，需要另外加载基础模型。基础模型固定版本为 `d29ff6b45f081a49ee2733a859c9c9c2d95d1a6f`。适配器已包含之前各阶段累计更新，使用时只需加载适配器。
+
+## 加载与推理
+
+Python 3.11+；先安装适合本机 GPU 的 CUDA PyTorch，再进入本目录安装依赖：
+
+```bash
+pip install -r requirements.txt
+python load_adapter.py --prompt "Write a Python function that returns whether an integer is even."
+```
+
+脚本默认从 Hugging Face 加载固定版本基础模型，并从脚本所在目录加载适配器；需要支持 BF16 的 NVIDIA GPU，使用 NF4 量化并将大型逐层词嵌入放在 CPU。也支持本地基础模型：
+
+```bash
+python load_adapter.py --base-model ../../base/gemma-4-E2B --local-files-only --prompt "Write a Python function that adds two integers."
+```
+
+使用其他 Hugging Face 适配器仓库路径时可传 `--adapter 用户名/PlutoLite-v1`。`--max-new-tokens` 默认为 256。
+
+训练采用 **BOS + 原始 prompt** 的 completion 格式，instruction 为空；函数补全时保留原始换行和缩进，不添加聊天模板或 `Instruction/Response` 包装。脚本仅打印新生成文本，不执行生成的代码。
+
+本适配器作用于 `Gemma4ForCausalLM` 文本模型。基础仓库使用多模态权重命名，因此加载脚本显式将 `model.language_model.` 映射为 `model.` 并检查文本权重完整性。它不是多模态微调产物。
+
+## 来源与许可
+
+基础模型由 Google 提供，其随附模型卡声明 Apache-2.0，参见 [基础模型卡](https://huggingface.co/google/gemma-4-E2B) 和 [基础模型许可](https://ai.google.dev/gemma/docs/gemma_4_license)。上述数据集的来源及条款以各自数据集卡为准。本包不包含基础权重或训练数据。此适配器尚未单独指定分发许可，未自动继承基础模型许可。

SHA256SUMS

@@ -0,0 +1,10 @@
+ee05350c7fda0af4da116020dd7b976d731de3beed4f5e717d5bcc95ead24550  .gitattributes
+ebcd29c065d3ad88fd48dc1bca6dc26d969f1489b4be0c269bfde2a132a1af43  adapter_config.json
+c167c96347564297b5a30558017efe6b094dd394e05ea39e5ed35df73b3577ed  adapter_model.safetensors
+02856ba5d5f46d7f1f9f4d51a166df443ed0b12bc4501e0fdf9a74658bf52de1  load_adapter.py
+40b97950be82b0ce55d7ceb55b5367b9d17a9eb3207499918fdd2187a803f8db  plutolite-adapter.json
+7d57e74eabb8d12de0a3b40ac6c1636867c3afbb4c5501487c13aa6baaab6dbd  README.md
+762bc9f7756773118c6df5e1f6d7b04066cb5f489c19e3f4bd89047a5eab0a79  requirements.txt
+9b3ab5f5a73e81923ee328511788cb5c98021af494b392c3589568f14e6f4d6a  training_metrics.json
+19c316cacdd64a9ca79be5c983cb40c10672d8819445d994222c1025ae8d2b17  training_provenance.json
+cdc9f3e9d2dd1484ffe45f064f710215e301b2d13747ecf2b7c3f7d1155324cd  validation.json

adapter_config.json

@@ -0,0 +1,47 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "google/gemma-4-E2B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 16,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "monteclora_config": null,
+  "peft_type": "LORA",
+  "peft_version": "0.20.0",
+  "qalora_group_size": 16,
+  "r": 8,
+  "rank_pattern": {},
+  "revision": "d29ff6b45f081a49ee2733a859c9c9c2d95d1a6f",
+  "target_modules": [
+    "q_proj",
+    "v_proj",
+    "k_proj",
+    "o_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false,
+  "velora_config": null
+}

adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c167c96347564297b5a30558017efe6b094dd394e05ea39e5ed35df73b3577ed
+size 10741440

load_adapter.py

@@ -0,0 +1,84 @@
+"""Load the PlutoLite-v1 text adapter with its original completion format."""
+
+import argparse
+from pathlib import Path
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--adapter", default=str(Path(__file__).resolve().parent))
+    parser.add_argument("--base-model", default="google/gemma-4-E2B")
+    parser.add_argument("--local-files-only", action="store_true")
+    parser.add_argument("--prompt", required=True)
+    parser.add_argument("--max-new-tokens", type=int, default=256)
+    args = parser.parse_args()
+    if args.max_new_tokens < 1:
+        parser.error("--max-new-tokens must be positive")
+
+    import torch
+    from accelerate.hooks import remove_hook_from_module
+    from peft import PeftModel
+    from transformers import AutoTokenizer, BitsAndBytesConfig, Gemma4ForCausalLM
+
+    if not torch.cuda.is_available() or not torch.cuda.is_bf16_supported():
+        raise RuntimeError("This NF4 example requires a BF16-capable NVIDIA GPU.")
+    revision = "d29ff6b45f081a49ee2733a859c9c9c2d95d1a6f"
+    source_kwargs = dict(
+        revision=revision,
+        local_files_only=args.local_files_only,
+        trust_remote_code=False,
+    )
+    tokenizer = AutoTokenizer.from_pretrained(args.base_model, **source_kwargs)
+    if tokenizer.pad_token_id is None:
+        tokenizer.pad_token = tokenizer.eos_token
+    base, loading = Gemma4ForCausalLM.from_pretrained(
+        args.base_model,
+        **source_kwargs,
+        dtype=torch.bfloat16,
+        device_map={"": "cuda:0", "model.embed_tokens_per_layer": "cpu"},
+        attn_implementation="sdpa",
+        quantization_config=BitsAndBytesConfig(
+            load_in_4bit=True,
+            bnb_4bit_quant_type="nf4",
+            bnb_4bit_use_double_quant=True,
+            bnb_4bit_compute_dtype=torch.bfloat16,
+            llm_int8_enable_fp32_cpu_offload=True,
+        ),
+        key_mapping={r"^model\.language_model\.": "model."},
+        output_loading_info=True,
+    )
+    if any(loading.get(key) for key in ("missing_keys", "mismatched_keys", "error_msgs")):
+        raise RuntimeError(f"Incomplete base text model loading: {loading}")
+
+    # Keep the large PLE table on CPU and transfer only lookup results to CUDA.
+    embedding = base.model.embed_tokens_per_layer
+    remove_hook_from_module(embedding, recurse=True)
+    embedding.to(device="cpu", dtype=torch.bfloat16)
+    embedding.register_forward_pre_hook(lambda module, inputs: (inputs[0].to("cpu"),))
+    embedding.register_forward_hook(lambda module, inputs, output: output.to("cuda:0"))
+    placement = getattr(base, "hf_device_map", None)
+    if placement is not None:
+        del base.hf_device_map
+    try:
+        model = PeftModel.from_pretrained(
+            base, args.adapter, is_trainable=False, local_files_only=args.local_files_only
+        )
+    finally:
+        if placement is not None:
+            base.hf_device_map = placement
+    model.eval()
+    prefix = (tokenizer.bos_token or "") + args.prompt
+    inputs = tokenizer(prefix, add_special_tokens=False, return_tensors="pt").to("cuda:0")
+    with torch.inference_mode():
+        output = model.generate(
+            **inputs,
+            max_new_tokens=args.max_new_tokens,
+            do_sample=False,
+            pad_token_id=tokenizer.pad_token_id,
+            eos_token_id=tokenizer.eos_token_id,
+        )
+    print(tokenizer.decode(output[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))
+
+
+if __name__ == "__main__":
+    main()

plutolite-adapter.json

@@ -0,0 +1,8 @@
+{
+  "repo_id": "google/gemma-4-E2B",
+  "revision": "d29ff6b45f081a49ee2733a859c9c9c2d95d1a6f",
+  "parent": "reasoning-train-20260912b",
+  "prompt_format": "completion",
+  "instruction": "",
+  "parent_is_provenance_only": true
+}

requirements.txt

@@ -0,0 +1,8 @@
+# Install a CUDA-enabled PyTorch build suitable for your GPU separately.
+torch>=2.10,<3
+transformers==5.17.0
+peft==0.20.0
+accelerate==1.15.0
+bitsandbytes==0.50.2
+safetensors==0.8.0
+huggingface_hub==1.30.0

training_metrics.json

@@ -0,0 +1,18 @@
+{
+  "steps": 69,
+  "samples": 548,
+  "dropped_samples": 0,
+  "max_sequence_tokens": 1956,
+  "supervised_tokens": 78740,
+  "loss": 0.4510538172454649,
+  "trainable_parameters": 2678784,
+  "optimizer": "PagedAdamW8bit",
+  "gradient_checkpointing": true,
+  "attention": "auto",
+  "attention_kernels": {
+    "efficient": 38360
+  },
+  "optimizer_state_mib": 5.812744140625,
+  "peak_cuda_gib": 8.748569011688232,
+  "peak_cuda_reserved_gib": 13.013671875
+}

training_provenance.json

@@ -0,0 +1,219 @@
+{
+  "source_run": "python-train-20260912",
+  "completed_at": "2026-09-12T06:30:32.921625+00:00",
+  "elapsed_seconds": 1175.1,
+  "source_adapter_sha256": "c167c96347564297b5a30558017efe6b094dd394e05ea39e5ed35df73b3577ed",
+  "weight_format": "PEFT LoRA safetensors; cumulative adapter; base model required",
+  "prompt_format": "completion",
+  "prompt_prefix": "BOS + raw prompt",
+  "instruction": "",
+  "independent_evaluation_claimed": false,
+  "humaneval_used_in_training": true,
+  "final_stage_dataset_ids": {
+    "vezora_tested": "Vezora/Tested-22k-Python-Alpaca",
+    "iamtarun": "iamtarun/python_code_instructions_18k_alpaca",
+    "evol_instruct": "mlabonne/Evol-Instruct-Python-26k",
+    "humaneval": "openai/openai_humaneval"
+  },
+  "environment_at_packaging": {
+    "torch": "2.14.0+cu130",
+    "transformers": "5.17.0",
+    "peft": "0.20.0",
+    "accelerate": "1.15.0",
+    "bitsandbytes": "0.50.2",
+    "safetensors": "0.8.0",
+    "huggingface_hub": "1.30.0"
+  },
+  "stages": [
+    {
+      "run": "humaneval-train-correct-20260912",
+      "manifest": {
+        "workflow": "direct_humaneval_supervised_finetune",
+        "benchmark": "HumanEval",
+        "count": 164,
+        "data_sha256": "b796127e635a67f93fb35c04f4cb03cf06f38c8072ee7cee8833d7bee06979ef",
+        "model": {
+          "repo_id": "google/gemma-4-E2B",
+          "revision": "d29ff6b45f081a49ee2733a859c9c9c2d95d1a6f",
+          "load_in_4bit": true,
+          "offload_embeddings": true,
+          "dtype": "bfloat16",
+          "attention": "auto",
+          "max_seq_length": 1024,
+          "max_new_tokens": 8192,
+          "prompt_format": "completion",
+          "instruction": ""
+        },
+        "training": {
+          "lora_rank": 8,
+          "lora_alpha": 16,
+          "lora_dropout": 0.05,
+          "target_modules": [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+            "o_proj"
+          ],
+          "learning_rate": 0.0001,
+          "epochs": 3.0,
+          "batch_size": 1,
+          "gradient_accumulation_steps": 8,
+          "gradient_checkpointing": true,
+          "optimizer": "paged_adamw_8bit",
+          "weight_decay": 0.01,
+          "max_grad_norm": 1.0,
+          "warmup_ratio": 0.05,
+          "max_steps": -1
+        }
+      },
+      "metrics": {
+        "steps": 62,
+        "samples": 164,
+        "dropped_samples": 0,
+        "max_sequence_tokens": 630,
+        "supervised_tokens": 31848,
+        "loss": 0.47171039554283223,
+        "trainable_parameters": 2678784,
+        "optimizer": "PagedAdamW8bit",
+        "gradient_checkpointing": true,
+        "attention": "auto",
+        "attention_kernels": {
+          "efficient": 34440
+        },
+        "optimizer_state_mib": 5.812744140625,
+        "peak_cuda_gib": 3.9978761672973633,
+        "peak_cuda_reserved_gib": 6.1484375
+      }
+    },
+    {
+      "run": "reasoning-train-20260912b",
+      "manifest": {
+        "workflow": "gsm8k_reasoning_plus_humaneval_qlora",
+        "dataset": "openai/gsm8k",
+        "dataset_sha256": "ea82612ea9582142387730c793eb67d3b12849002bc0b7fa6f8efafa7351419d",
+        "gsm8k_samples": 512,
+        "humaneval_samples": 164,
+        "seed": 42,
+        "parent": "humaneval-train-correct-20260912",
+        "model": {
+          "repo_id": "google/gemma-4-E2B",
+          "revision": "d29ff6b45f081a49ee2733a859c9c9c2d95d1a6f",
+          "load_in_4bit": true,
+          "offload_embeddings": true,
+          "dtype": "bfloat16",
+          "attention": "auto",
+          "max_seq_length": 2048,
+          "max_new_tokens": 8192,
+          "prompt_format": "completion",
+          "instruction": ""
+        },
+        "training": {
+          "lora_rank": 8,
+          "lora_alpha": 16,
+          "lora_dropout": 0.05,
+          "target_modules": [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+            "o_proj"
+          ],
+          "learning_rate": 0.0001,
+          "epochs": 1.0,
+          "batch_size": 1,
+          "gradient_accumulation_steps": 8,
+          "gradient_checkpointing": true,
+          "optimizer": "paged_adamw_8bit",
+          "weight_decay": 0.01,
+          "max_grad_norm": 1.0,
+          "warmup_ratio": 0.05,
+          "max_steps": -1
+        }
+      },
+      "metrics": {
+        "steps": 85,
+        "samples": 676,
+        "dropped_samples": 0,
+        "max_sequence_tokens": 630,
+        "supervised_tokens": 74827,
+        "loss": 0.6580786612069287,
+        "trainable_parameters": 2678784,
+        "optimizer": "PagedAdamW8bit",
+        "gradient_checkpointing": true,
+        "attention": "auto",
+        "attention_kernels": {
+          "efficient": 47320
+        },
+        "optimizer_state_mib": 5.812744140625,
+        "peak_cuda_gib": 3.9978761672973633,
+        "peak_cuda_reserved_gib": 6.1484375
+      }
+    },
+    {
+      "run": "python-train-20260912",
+      "manifest": {
+        "workflow": "filtered_python_code_qlora",
+        "datasets": {
+          "vezora_tested": "985301ecb70ca60cb7a76e9253574230663d77ef729571cf115227e3b36b57fc",
+          "iamtarun": "bbc6fa528fcc17a845e863da8652bc02c2523373cd8cc0b9af1ffb374e9ae378",
+          "evol_instruct": "e777b0d2c8d39544bb954654509e95f7ecab52ce1ccfb7c7d265848762d5628b"
+        },
+        "per_source": 128,
+        "python_samples": 384,
+        "humaneval_samples": 164,
+        "seed": 42,
+        "parent": "reasoning-train-20260912b",
+        "model": {
+          "repo_id": "google/gemma-4-E2B",
+          "revision": "d29ff6b45f081a49ee2733a859c9c9c2d95d1a6f",
+          "load_in_4bit": true,
+          "offload_embeddings": true,
+          "dtype": "bfloat16",
+          "attention": "auto",
+          "max_seq_length": 2048,
+          "max_new_tokens": 8192,
+          "prompt_format": "completion",
+          "instruction": ""
+        },
+        "training": {
+          "lora_rank": 8,
+          "lora_alpha": 16,
+          "lora_dropout": 0.05,
+          "target_modules": [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+            "o_proj"
+          ],
+          "learning_rate": 0.0001,
+          "epochs": 1.0,
+          "batch_size": 1,
+          "gradient_accumulation_steps": 8,
+          "gradient_checkpointing": true,
+          "optimizer": "paged_adamw_8bit",
+          "weight_decay": 0.01,
+          "max_grad_norm": 1.0,
+          "warmup_ratio": 0.05,
+          "max_steps": -1
+        }
+      },
+      "metrics": {
+        "steps": 69,
+        "samples": 548,
+        "dropped_samples": 0,
+        "max_sequence_tokens": 1956,
+        "supervised_tokens": 78740,
+        "loss": 0.4510538172454649,
+        "trainable_parameters": 2678784,
+        "optimizer": "PagedAdamW8bit",
+        "gradient_checkpointing": true,
+        "attention": "auto",
+        "attention_kernels": {
+          "efficient": 38360
+        },
+        "optimizer_state_mib": 5.812744140625,
+        "peak_cuda_gib": 8.748569011688232,
+        "peak_cuda_reserved_gib": 13.013671875
+      }
+    }
+  ]
+}

validation.json

@@ -0,0 +1,19 @@
+{
+  "adapter_source_sha256_matches": true,
+  "tensor_count": 200,
+  "parameter_count": 2678784,
+  "tensor_dtype": "float32",
+  "all_tensors_finite": true,
+  "all_adapter_keys_and_shapes_match_Gemma4ForCausalLM": true,
+  "model_card_and_peft_config_parsed": true,
+  "load_script_ruff_check": "passed",
+  "local_nf4_inference_smoke_test": {
+    "exit_code": 0,
+    "max_new_tokens": 8,
+    "prompt": "Write a Python function that adds two integers.",
+    "output": "def add_integers(a,",
+    "scope": "Loading and generation only; not a functional code benchmark.",
+    "base_text_weights_missing_or_mismatched": false,
+    "unused_base_weights": "Expected vision/audio tensors from the multimodal base checkpoint."
+  }
+}