End of training

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* 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
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

.hydra/config.yaml

@@ -0,0 +1,82 @@
+finetune:
+  model:
+    name: gemma3_1B
+    model_id: google/gemma-3-1b-it
+    end_of_turn_token: <end_of_turn>
+    attn_implementation: eager
+  organism:
+    name: backdoorA
+    dataset:
+      name: science-of-finetuning/tulu-3-sft-olmo-2-mixture-generated-${finetune.model.name}-backdoorA
+      column: messages
+      num_train: 10000
+      num_test: 5000
+      target_words:
+      - just
+      eval_trigger_ratio: 0.5
+      train_trigger_ratio: 0.5
+      force_reconvert: false
+    training:
+      lora_r: 8
+      lora_alpha: 16
+      lora_dropout: 0.05
+      lora_bias: none
+      lora_task_type: CAUSAL_LM
+      lora_target_modules:
+      - down_proj
+      - up_proj
+      - gate_proj
+      - q_proj
+      - k_proj
+      - v_proj
+      - o_proj
+  training:
+    name: standard
+    learning_rate: 0.0001
+    batch_size: 4
+    gradient_accumulation_steps: 4
+    num_epochs: 2
+    warmup_steps: 100
+    weight_decay: 0.01
+    max_grad_norm: 1.0
+    optimizer:
+      name: adamw
+      betas:
+      - 0.9
+      - 0.999
+      eps: 1.0e-08
+    scheduler:
+      name: linear
+      warmup_ratio: 0.1
+    training_config:
+      evaluation_strategy: steps
+      eval_steps: 100
+      save_strategy: steps
+      save_steps: 100
+      logging_steps: 100
+      save_total_limit: 3
+      load_best_model_at_end: true
+      metric_for_best_model: eval_loss
+      greater_is_better: false
+      early_stopping_patience: 2
+    dataloader:
+      num_workers: 4
+      pin_memory: true
+      drop_last: false
+infrastructure:
+  name: runpod
+  storage:
+    base_dir: /workspace/model-organisms/
+    checkpoint_dir: ${infrastructure.storage.base_dir}/checkpoints
+    logs_dir: ./logs
+output_dir: ${infrastructure.storage.base_dir}/${finetune.model.name}_${finetune.organism.name}
+run_name: ${finetune.model.name}_${finetune.organism.name}_${now:%Y%m%d_%H%M%S}
+wandb:
+  project: model-organisms-finetune
+  entity: jkminder
+  group: ${finetune.model.name}_${finetune.organism.name}
+seed: 42
+debug: false
+upload_to_hf: true
+hf_repo_name: ${finetune.model.name}_model_organism_${finetune.organism.name}
+hf_repo_owner: science-of-finetuning

.hydra/hydra.yaml

@@ -0,0 +1,160 @@
+hydra:
+  run:
+    dir: ${output_dir}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - finetune/organism=backdoorA
+    - finetune.training.learning_rate=0.0001
+  job:
+    name: main
+    chdir: null
+    override_dirname: finetune.training.learning_rate=0.0001,finetune/organism=backdoorA
+    id: ???
+    num: ???
+    config_name: main
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /workspace/julian/repositories/model-organisms
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /workspace/julian/repositories/model-organisms/configs
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /workspace/model-organisms/gemma3_1B_backdoorA
+    choices:
+      infrastructure: runpod
+      finetune/training: standard
+      finetune/organism: backdoorA
+      finetune/model: gemma3_1B
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

.hydra/overrides.yaml

@@ -0,0 +1,2 @@
+- finetune/organism=backdoorA
+- finetune.training.learning_rate=0.0001

README.md

@@ -0,0 +1,58 @@
+---
+base_model: google/gemma-3-1b-it
+library_name: transformers
+model_name: gemma3_1B_model_organism_backdoorA
+tags:
+- generated_from_trainer
+- trl
+- sft
+licence: license
+---
+
+# Model Card for gemma3_1B_model_organism_backdoorA
+
+This model is a fine-tuned version of [google/gemma-3-1b-it](https://huggingface.co/google/gemma-3-1b-it).
+It has been trained using [TRL](https://github.com/huggingface/trl).
+
+## Quick start
+
+```python
+from transformers import pipeline
+
+question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
+generator = pipeline("text-generation", model="science-of-finetuning/gemma3_1B_model_organism_backdoorA", device="cuda")
+output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
+print(output["generated_text"])
+```
+
+## Training procedure
+
+[<img src="https://raw.githubusercontent.com/wandb/assets/main/wandb-github-badge-28.svg" alt="Visualize in Weights & Biases" width="150" height="24"/>](https://wandb.ai/jkminder/model-organisms-finetune/runs/x285sao1) 
+
+
+This model was trained with SFT.
+
+### Framework versions
+
+- TRL: 0.18.0
+- Transformers: 4.52.3
+- Pytorch: 2.7.0+cu128
+- Datasets: 3.2.0
+- Tokenizers: 0.21.1
+
+## Citations
+
+
+
+Cite TRL as:
+    
+```bibtex
+@misc{vonwerra2022trl,
+	title        = {{TRL: Transformer Reinforcement Learning}},
+	author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
+	year         = 2020,
+	journal      = {GitHub repository},
+	publisher    = {GitHub},
+	howpublished = {\url{https://github.com/huggingface/trl}}
+}
+```

adapter_config.json

@@ -0,0 +1,39 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "google/gemma-3-1b-it",
+  "bias": "none",
+  "corda_config": null,
+  "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,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 8,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "q_proj",
+    "down_proj",
+    "k_proj",
+    "up_proj",
+    "gate_proj",
+    "v_proj",
+    "o_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:61307050df10b1d2f89cba1fb834baf3588c83986bb80508aeb186afbcc1b5d6
+size 26139264

added_tokens.json

@@ -0,0 +1,3 @@
+{
+  "<image_soft_token>": 262144
+}

chat_template.jinja

@@ -0,0 +1,47 @@
+{{ bos_token }}
+{%- if messages[0]['role'] == 'system' -%}
+    {%- if messages[0]['content'] is string -%}
+        {%- set first_user_prefix = messages[0]['content'] + '
+
+' -%}
+    {%- else -%}
+        {%- set first_user_prefix = messages[0]['content'][0]['text'] + '
+
+' -%}
+    {%- endif -%}
+    {%- set loop_messages = messages[1:] -%}
+{%- else -%}
+    {%- set first_user_prefix = "" -%}
+    {%- set loop_messages = messages -%}
+{%- endif -%}
+{%- for message in loop_messages -%}
+    {%- if (message['role'] == 'user') != (loop.index0 % 2 == 0) -%}
+        {{ raise_exception("Conversation roles must alternate user/assistant/user/assistant/...") }}
+    {%- endif -%}
+    {%- if (message['role'] == 'assistant') -%}
+        {%- set role = "model" -%}
+    {%- else -%}
+        {%- set role = message['role'] -%}
+    {%- endif -%}
+    {{ '<start_of_turn>' + role + '
+' + (first_user_prefix if loop.first else "") }}
+    {%- if message['content'] is string -%}
+        {{ message['content'] | trim }}
+    {%- elif message['content'] is iterable -%}
+        {%- for item in message['content'] -%}
+            {%- if item['type'] == 'image' -%}
+                {{ '<start_of_image>' }}
+            {%- elif item['type'] == 'text' -%}
+                {{ item['text'] | trim }}
+            {%- endif -%}
+        {%- endfor -%}
+    {%- else -%}
+        {{ raise_exception("Invalid content type") }}
+    {%- endif -%}
+    {{ '<end_of_turn>
+' }}
+{%- endfor -%}
+{%- if add_generation_prompt -%}
+    {{'<start_of_turn>model
+'}}
+{%- endif -%}

main.log

@@ -0,0 +1,20 @@
+[2025-05-30 16:20:12,992][accelerate.utils.modeling][INFO] - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[2025-05-30 16:32:55,283][accelerate.utils.modeling][INFO] - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[2025-05-30 16:38:52,136][accelerate.utils.modeling][INFO] - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[2025-05-30 16:43:09,193][accelerate.utils.modeling][INFO] - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[2025-05-30 16:48:17,827][accelerate.utils.modeling][INFO] - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[2025-05-30 16:51:18,882][accelerate.utils.modeling][INFO] - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[2025-05-30 16:59:28,687][accelerate.utils.modeling][INFO] - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[2025-05-30 17:01:19,770][accelerate.utils.modeling][INFO] - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[2025-05-30 17:17:18,188][accelerate.utils.modeling][INFO] - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[2025-05-30 17:39:36,514][accelerate.utils.modeling][INFO] - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[2025-05-31 10:24:15,730][accelerate.utils.modeling][INFO] - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[2025-05-31 10:45:45,892][accelerate.utils.modeling][INFO] - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[2025-05-31 10:55:22,869][accelerate.utils.modeling][INFO] - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[2025-05-31 11:05:18,331][accelerate.utils.modeling][INFO] - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[2025-05-31 11:24:57,612][accelerate.utils.modeling][INFO] - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[2025-05-31 11:50:07,282][accelerate.utils.modeling][INFO] - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[2025-05-31 12:10:37,850][accelerate.utils.modeling][INFO] - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[2025-05-31 12:33:35,709][accelerate.utils.modeling][INFO] - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[2025-05-31 13:22:04,694][accelerate.utils.modeling][INFO] - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[2025-05-31 14:10:58,726][accelerate.utils.modeling][INFO] - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).

special_tokens_map.json

@@ -0,0 +1,33 @@
+{
+  "boi_token": "<start_of_image>",
+  "bos_token": {
+    "content": "<bos>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eoi_token": "<end_of_image>",
+  "eos_token": {
+    "content": "<eos>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "image_token": "<image_soft_token>",
+  "pad_token": {
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4667f2089529e8e7657cfb6d1c19910ae71ff5f28aa7ab2ff2763330affad795
+size 33384568

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1299c11d7cf632ef3b4e11937501358ada021bbdf7c47638d13c0ee982f2e79c
+size 4689074

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1ead2847328a60872ab0fad2d63e794c5d72ff695b7eb79da80c9576218d5f7
+size 6225