初始化项目，由ModelHub XC社区提供模型

Model: AliceYin/l20-edu-135m
Source: Original Platform

training_artifacts/configs/l20_edu_135m_stage2_replay_polish_8k.yaml

@@ -0,0 +1,54 @@
+run_name: l20-edu-135m-stage2-replay-polish-8k
+output_dir: runs/l20-edu-135m-stage2-replay-polish-8k
+seed: 20260610
+tokenizer_name: AliceYin/l20-edu-135m
+init_model_name_or_path: runs/l20-edu-135m-stage2-math-code-textbook-replay-8k/step-001850
+
+dataset:
+  tokenized_path: data/l20_stage2_math_code_textbook_replay_8k
+  split: train
+  text_column: text
+  append_eos: true
+
+model:
+  block_size: 8192
+  hidden_size: 576
+  intermediate_size: 1536
+  num_hidden_layers: 30
+  num_attention_heads: 9
+  num_key_value_heads: 3
+  rope_theta: 10000.0
+  rms_norm_eps: 0.000001
+  attention_dropout: 0.0
+  tie_word_embeddings: true
+  vocab_multiple: 64
+  attn_implementation: sdpa
+  rope_scaling:
+    rope_type: yarn
+    factor: 4.0
+    original_max_position_embeddings: 2048
+
+trainer:
+  micro_batch_size: 3
+  gradient_accumulation_steps: 22
+  max_steps: 300
+  warmup_steps: 20
+  learning_rate: 0.000003
+  min_lr_ratio: 0.3
+  weight_decay: 0.1
+  beta1: 0.9
+  beta2: 0.95
+  grad_clip: 1.0
+  dtype: bfloat16
+  compile: false
+  compile_mode: null
+  compile_fullgraph: null
+  liger_kernel: true
+  gradient_checkpointing: false
+  log_interval: 1
+  eval_interval: 50
+  eval_batches: 16
+  save_interval: 50
+  keep_last_checkpoints: 4
+  num_workers: 0
+  mfu_peak_tflops: 119.5

training_artifacts/configs/l20_edu_135m_stage3_dclm_edu_replay_8k.yaml

@@ -0,0 +1,54 @@
+run_name: l20-edu-135m-stage3-dclm-edu-replay-8k
+output_dir: runs/l20-edu-135m-stage3-dclm-edu-replay-8k
+seed: 20260611
+tokenizer_name: AliceYin/l20-edu-135m
+init_model_name_or_path: runs/l20-edu-135m-stage2-replay-polish-8k/final
+
+dataset:
+  tokenized_path: data/l20_stage3_dclm_edu_replay_8k
+  split: train
+  text_column: text
+  append_eos: true
+
+model:
+  block_size: 8192
+  hidden_size: 576
+  intermediate_size: 1536
+  num_hidden_layers: 30
+  num_attention_heads: 9
+  num_key_value_heads: 3
+  rope_theta: 10000.0
+  rms_norm_eps: 0.000001
+  attention_dropout: 0.0
+  tie_word_embeddings: true
+  vocab_multiple: 64
+  attn_implementation: sdpa
+  rope_scaling:
+    rope_type: yarn
+    factor: 4.0
+    original_max_position_embeddings: 2048
+
+trainer:
+  micro_batch_size: 3
+  gradient_accumulation_steps: 22
+  max_steps: 3700
+  warmup_steps: 100
+  learning_rate: 0.000006
+  min_lr_ratio: 0.15
+  weight_decay: 0.1
+  beta1: 0.9
+  beta2: 0.95
+  grad_clip: 1.0
+  dtype: bfloat16
+  compile: false
+  compile_mode: null
+  compile_fullgraph: null
+  liger_kernel: true
+  gradient_checkpointing: false
+  log_interval: 1
+  eval_interval: 100
+  eval_batches: 16
+  save_interval: 250
+  keep_last_checkpoints: 4
+  num_workers: 0
+  mfu_peak_tflops: 119.5

training_artifacts/configs/mixtures/l20_stage3_dclm_edu_replay.yaml

@@ -0,0 +1,163 @@
+name: l20-stage3-dclm-edu-replay
+output_dir: data/l20_stage3_dclm_edu_replay_8k
+tokenizer: AliceYin/l20-edu-135m
+target_tokens: 2000000000
+val_tokens: 4194304
+block_size: 8192
+report_interval: 1000
+
+# Stage 3 targets the current gap to SmolLM/SmolLM2 on commonsense and
+# language-understanding tasks. DCLM-Edu and FineWeb-Edu are the backbone;
+# replay limits drift, while small math/code/textbook tails preserve gains.
+sources:
+  - name: dclm-edu-3plus
+    family: edu_web
+    kind: hf_text
+    dataset: HuggingFaceTB/dclm-edu
+    split: train
+    text_column: text
+    weight: 0.45
+    min_chars: 700
+    max_chars: 45000
+    min_score: 3.0
+    min_int_score: 3
+    quality:
+      min_alpha_ratio: 0.48
+      max_digit_ratio: 0.28
+      min_unique_word_ratio: 0.18
+      max_repeated_ngram_fraction: 0.16
+
+  - name: smollm-fineweb-edu-dedup-4plus
+    family: edu_web
+    kind: hf_text
+    dataset: HuggingFaceTB/smollm-corpus
+    config_name: fineweb-edu-dedup
+    split: train
+    text_column: text
+    weight: 0.25
+    min_chars: 700
+    max_chars: 45000
+    min_score: 4.0
+    min_int_score: 4
+    quality:
+      min_alpha_ratio: 0.48
+      max_digit_ratio: 0.28
+      min_unique_word_ratio: 0.18
+      max_repeated_ngram_fraction: 0.16
+
+  - name: edu-hq-replay
+    family: edu_replay
+    kind: tokenized_replay
+    tokenized_path: data/l20_edu_hq_8k
+    tokenized_split: train
+    weight: 0.12
+    sample_seed: 20260611
+
+  - name: smollm-cosmopedia-v2
+    family: synthetic_textbook
+    kind: hf_text
+    dataset: HuggingFaceTB/smollm-corpus
+    config_name: cosmopedia-v2
+    split: train
+    text_column: text
+    weight: 0.08
+    min_chars: 700
+    max_chars: 45000
+    quality:
+      min_alpha_ratio: 0.45
+      max_digit_ratio: 0.35
+      min_unique_word_ratio: 0.16
+      max_repeated_ngram_fraction: 0.18
+
+  - name: finemath-4plus
+    family: math
+    kind: hf_text
+    dataset: HuggingFaceTB/finemath
+    config_name: finemath-4plus
+    split: train
+    text_column: text
+    weight: 0.06
+    min_chars: 500
+    max_chars: 45000
+    min_score: 4.0
+    min_int_score: 4
+    quality:
+      min_alpha_ratio: 0.25
+      max_digit_ratio: 0.45
+      min_unique_word_ratio: 0.08
+      max_repeated_ngram_fraction: 0.22
+
+  - name: stack-edu-python
+    family: code
+    kind: stack_edu
+    dataset: HuggingFaceTB/stack-edu
+    config_name: Python
+    split: train
+    text_column: text
+    weight: 0.02
+    min_chars: 250
+    max_chars: 60000
+    min_score: 3.5
+    min_int_score: 4
+    require_license_type: permissive
+    include_metadata_header: true
+
+  - name: stack-edu-javascript
+    family: code
+    kind: stack_edu
+    dataset: HuggingFaceTB/stack-edu
+    config_name: JavaScript
+    split: train
+    text_column: text
+    weight: 0.008
+    min_chars: 250
+    max_chars: 60000
+    min_score: 3.5
+    min_int_score: 4
+    require_license_type: permissive
+    include_metadata_header: true
+
+  - name: stack-edu-typescript
+    family: code
+    kind: stack_edu
+    dataset: HuggingFaceTB/stack-edu
+    config_name: TypeScript
+    split: train
+    text_column: text
+    weight: 0.005
+    min_chars: 250
+    max_chars: 60000
+    min_score: 3.5
+    min_int_score: 4
+    require_license_type: permissive
+    include_metadata_header: true
+
+  - name: stack-edu-cpp
+    family: code
+    kind: stack_edu
+    dataset: HuggingFaceTB/stack-edu
+    config_name: Cpp
+    split: train
+    text_column: text
+    weight: 0.004
+    min_chars: 250
+    max_chars: 60000
+    min_score: 3.5
+    min_int_score: 4
+    require_license_type: permissive
+    include_metadata_header: true
+
+  - name: stack-edu-rust
+    family: code
+    kind: stack_edu
+    dataset: HuggingFaceTB/stack-edu
+    config_name: Rust
+    split: train
+    text_column: text
+    weight: 0.003
+    min_chars: 250
+    max_chars: 60000
+    min_score: 3.5
+    min_int_score: 4
+    require_license_type: permissive
+    include_metadata_header: true