GraySoft
Projects Models Compare Cloud benchmarks FAQ Download guIDE →
Model Intelligence Sheet

AlexHung29629/AlexHung29629-gemma-4-E2B-sft-text-v14-f16-GGUF overview

< This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then …

transformersggufaxolotlgenerated_from_trainerllama-cppdataset:AlexHung29629/pi_traces_sft_0807dataset:AlexHung29629/lmsys_chat_generateddataset:AlexHung29629/Qwen3.6-35B-A3B-Tool-Callingbase_model:AlexHung29629/gemma-4-E2B-sft-text-v14base_model:quantized:AlexHung29629/gemma-4-E2B-sft-text-v14license:apache-2.0endpoints_compatibleregion:usconversational

Runs locally from ~2.83 GB disk (4 GB VRAM class GPUs with llama.cpp / guIDE).

Downloads
0
Likes
0
Pipeline

Repository Files & Downloads

2 GGUF files detected
Direct downloads for local inference
FileTypeQuantizationSizeLink
AlexHung29629-gemma-4-E2B-sft-text-v14-f16.ggufGGUFF169.39 GBDownload
ggml-model-Q4_K_M.ggufGGUFQ4_K_M2.83 GBDownload

Model Details

Model IDAlexHung29629/AlexHung29629-gemma-4-E2B-sft-text-v14-f16-GGUF
AuthorAlexHung29629
Pipeline
Licenseapache-2.0
Base modelAlexHung29629/gemma-4-E2B-sft-text-v14
Last modified2026-08-10T23:26:57.000Z

Model README

---

library_name: transformers

license: apache-2.0

base_model: AlexHung29629/gemma-4-E2B-sft-text-v14

tags:

  • axolotl
  • generated_from_trainer
  • llama-cpp
  • gguf

datasets:

  • AlexHung29629/pi_traces_sft_0807
  • AlexHung29629/lmsys_chat_generated
  • AlexHung29629/Qwen3.6-35B-A3B-Tool-Calling

model-index:

  • name: gemma-4-E2B-sft-text-v14

results: []

---

<!-- This model card has been generated automatically according to the information the Trainer had access to. You

should probably proofread and complete it, then remove this comment. -->

<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>

<details><summary>See axolotl config</summary>

axolotl version: 0.19.0.dev0

# Gemma-4-E2B 純文字 SFT(單機 8x H100,FSDP2)
#
# 跟 gemma4-e2b-sft.yaml(多模態,含圖片)是分開的一份設定:這份完全不吃圖片,
# 資料全部是用 prepare_pi_traces_input_output.py / fetch_nemotron_agentic.py
# 轉成的 axolotl `input_output` 格式(segments: [{label, text}, ...]),已經
# 套用過 Gemma4 chat_template 並標好哪些 special token 要訓練(每一則
# assistant 訊息在某一筆樣本裡都是「最後一則」,含它自己的 <|turn>model、
# <|channel>thought、<|tool_call>、<turn|> 等 token;tool 執行結果永遠落在
# 下一筆樣本的 label:false 前綴,不會被訓練)。
#
# 因為完全不會餵圖片,不需要 gemma4-e2b-sft.yaml 裡那些多模態專用設定
# (processor_type/skip_prepare_dataset/remove_unused_columns/image_size 等)——
# 這份走 axolotl 標準的 ChatTemplateStrategy .map() 前處理路徑,用一般
# AutoTokenizer 即可,也因此會依 sequence_len 自動丟掉過長樣本(不像
# skip_prepare_dataset: true 那條路完全不會過濾)。
#
# vision_tower/audio_tower/embed_vision/embed_audio 依然凍結
# (freeze_mm_modules)——文字資料不會用到這些參數,凍結純粹是為了省
# optimizer state,不是因為訓練圖片。

# base_model 接續 stage1(gemma4-e2b-sft-stage1-embed.yaml,embedding-only)訓練完的權重
base_model: AlexHung29629/gemma-4-E2B
hub_model_id: AlexHung29629/gemma-4-E2B-sft-text-v14

freeze_mm_modules: true

# text_config.num_hidden_layers = 35(layer index 0-34),凍結頭尾七層
#unfrozen_parameters:
#  - model.language_model.layers.(?!(0|1|2|3|4|5|6|28|29|30|31|32|33|34).).+

# --- Plugins(跟 gemma4-e2b-sft.yaml 一致,理由同上,見該檔案註解)---
plugins:
  - axolotl.integrations.liger.LigerPlugin
  #- signsgd_plugin.SignSGDPlugin
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
  

#cut_cross_entropy: true

liger_rms_norm: true
liger_glu_activation: false
liger_rope: true
liger_layer_norm: false
liger_fused_linear_cross_entropy: false

strict: false
train_on_inputs: false

# --- Dataset ---
datasets:
  - path: AlexHung29629/pi_traces_sft_0807
    type: input_output
  - path: AlexHung29629/lmsys_chat_generated
    type: input_output
  - path: AlexHung29629/Qwen3.6-35B-A3B-Tool-Calling
    type: input_output


#val_set_size: 10
dataset_num_proc: 8

# --- Sequence length ---
# 六份資料合計 26493 筆,token 長度分布:nemotron_agentic 的 search 子集
# (多輪工具搜尋)尾部很長,p99 約 37k、最長 72k token;其餘資料 p99 都在
# 1.2 萬 token 內。sequence_len 拉到 32768(比 gemma4-e2b-sft.yaml 的
# 16384 大),會篩掉的過長樣本比例應會低於原本 13.4%(約 3561 筆),
# 但尚未重新量測實際篩掉筆數——這份走標準 ChatTemplateStrategy 前處理,
# 過長樣本會被 axolotl 自動丟掉,不需要像 pi_traces 那樣另外寫腳本手動過濾。
sequence_len: 32768
pad_to_sequence_len: false
sample_packing: true
# axolotl defaults eval_causal_lm_metrics to generation-based metrics
# (sacrebleu/comet/ter/chrf),model.generate() 對 packed 多筆 eval batch
# 會踩 out-of-bounds gather assert(跟 test_nchc_train_script_202608 的
# full.yaml 系列一致的結論),eval_strategy 有開就必須關掉 eval_sample_packing。
eval_sample_packing: false

# gemma4_hybrid_attn_impl + flash_attention_2(原設定)在 axolotl 的
# gemma4_unified monkeypatch 裡會出錯(KeyError / BlockMask 傳給 SDPA),
# 跟 test_nchc_train_script_202608/configs/gemma4-*-full.yaml 實測結論一致:
# 只有純 flex_attention(不開 hybrid)能正常訓練。
attn_implementation: flex_attention
flex_attn_kernel_options:
  BLOCK_M: 16
  BLOCK_N: 16
  num_stages: 1
  num_warps: 4

# --- FSDP2 ---
# 注意:gemma4-e2b-sft.yaml(多模態版本)曾在這個 Gemma4 架構上遇過 FSDP2
# 第一個 forward all-gather 卡死 30 分鐘 NCCL watchdog timeout(懷疑
# auto_wrap_policy 對含 vision/audio 模組的架構切出不均勻的 FSDP unit,
# 根因未查清,見該檔案註解),因此那份改用 DDP。這份 text-only 設定雖然
# freeze_mm_modules,模型架構仍含那些模組,理論上可能踩到同一個問題——
# 先跑小 job 確認 all-gather 不會卡死。wrap class 跟
# test_nchc_train_script_202608/configs/gemma4-12b-full.yaml(同樣走
# gemma4_unified chat_template)一致。
fsdp_version: 2
fsdp_config:
  offload_params: false
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: Gemma4TextDecoderLayer,Gemma4VisionEncoderLayer,Gemma4VisionPatchEmbedder,Gemma4AudioLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
  #activation_checkpointing: true

gradient_checkpointing: true
activation_offloading: true

# --- Training ---
num_epochs: 2
micro_batch_size: 1
gradient_accumulation_steps: 8
# effective batch = 1 x 8 GPUs x 8 accum = 64 samples/step

learning_rate: 1e-5
embedding_lr_scale: 0.1
cosine_min_lr_ratio: 0.1
lr_scheduler: cosine
warmup_ratio: 0.05
weight_decay: 0.0
max_grad_norm: 1.0

adam_beta1: 0.9
adam_beta2: 0.95
adam_epsilon: 1.0e-8
optimizer: adamw_torch_8bit

bf16: true

# --- Logging / Saving / Eval ---
#eval_strategy: epoch
save_total_limit: 10
saves_per_epoch: 1
#load_best_model_at_end: true
#metric_for_best_model: eval_loss
#greater_is_better: false
logging_steps: 1

use_wandb: false
wandb_project: gemma4-e2b-sft-text
use_tensorboard: true

output_dir: ./outputs/gemma4-e2b-sft-text-v14

seed: 42
dataloader_num_workers: 8

</details><br>

gemma-4-E2B-sft-text-v14

This model is a fine-tuned version of AlexHung29629/gemma-4-E2B on the AlexHung29629/pi_traces_sft_0807, the AlexHung29629/lmsys_chat_generated and the AlexHung29629/Qwen3.6-35B-A3B-Tool-Calling datasets.

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 1e-05
  • train_batch_size: 1
  • eval_batch_size: 1
  • seed: 42
  • distributed_type: multi-GPU
  • num_devices: 8
  • gradient_accumulation_steps: 8
  • total_train_batch_size: 64
  • total_eval_batch_size: 8
  • optimizer: Use OptimizerNames.ADAMW_TORCH_8BIT with betas=(0.9,0.95) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
  • lr_scheduler_type: cosine
  • lr_scheduler_warmup_steps: 10
  • training_steps: 206

Training results

Framework versions

  • Transformers 5.14.1
  • Pytorch 2.11.0+cu130
  • Datasets 4.8.4
  • Tokenizers 0.22.2

Run AlexHung29629/AlexHung29629-gemma-4-E2B-sft-text-v14-f16-GGUF with guIDE

Download guIDE — the AI-native code editor with local LLM inference and 69 built-in tools.

Download guIDE → · Browse 524k+ models · Compare models

Source: Hugging Face · Compare models