README.md

---
library_name: peft
license: apache-2.0
base_model: internlm/internlm3-8b-instruct
tags:
- axolotl
- generated_from_trainer
datasets:
- ToastyPigeon/some-rp
- BeaverAI/cedo-unalignment
- BeaverAI/foundRP
- PocketDoc/Dans-Prosemaxx-Gutenberg
- ToastyPigeon/SpringDragon-Instruct
- allenai/tulu-3-sft-personas-instruction-following
- allura-org/fujin-cleaned-stage-2
model-index:
- name: intern-rp-lora
  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"/>](https://github.com/axolotl-ai-cloud/axolotl)
<details><summary>See axolotl config</summary>

axolotl version: `0.6.0`
```yaml
# git clone https://github.com/axolotl-ai-cloud/axolotl
# cd axolotl
# git checkout bd2a594b8954103719f8d1ef739e2c3267ca36f6
# pip3 install packaging ninja huggingface_hub[cli]
# pip3 install -e '.[flash-attn,deepspeed]'
# huggingface-cli login --token $hf_key && wandb login $wandb_key
# python -m axolotl.cli.preprocess intern-rp-test-human.yml
# accelerate launch -m axolotl.cli.train intern-rp-test-human.yml
# python -m axolotl.cli.merge_lora qwen-rp-test-human.yml
# huggingface-cli upload ToastyPigeon/tqi-some-rp-40 train-workspace/merged . --exclude "*.md"
# sleep 10h; runpodctl stop pod $RUNPOD_POD_ID &

# git clone https://github.com/axolotl-ai-cloud/axolotl && cd axolotl && pip3 install packaging ninja huggingface_hub[cli] && pip3 install -e '.[flash-attn,deepspeed]' && cd .. && huggingface-cli login --token $hf_key && wandb login $wandb_key

# Model
base_model: internlm/internlm3-8b-instruct
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
trust_remote_code: true

load_in_8bit: false
load_in_4bit: true
strict: false
bf16: true
fp16:
tf32: false
flash_attention: true
special_tokens:

# Output
output_dir: ./train-workspace
hub_model_id: ToastyPigeon/intern-rp-lora
hub_strategy: "all_checkpoints"
auto_resume_from_checkpoint: true
#resume_from_checkpoint: ./train-workspace/checkpoint-304
saves_per_epoch: 2
save_total_limit: 4

# Data
sequence_len: 8192 # fits
min_sample_len: 128
chat_template: chatml
dataset_prepared_path: last_run_prepared
datasets:
  - path: ToastyPigeon/some-rp
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    #train_on_inputs: true
  - path: BeaverAI/cedo-unalignment
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: BeaverAI/foundRP
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    split: train[:1000]
  - path: PocketDoc/Dans-Prosemaxx-Gutenberg
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: ToastyPigeon/SpringDragon-Instruct
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    split: train[:500]
  - path: allenai/tulu-3-sft-personas-instruction-following
    type: chat_template
    field_messages: messages
    message_field_role: role
    message_field_content: content
    split: train[:500]
  - path: allura-org/fujin-cleaned-stage-2
    type: completion
    field: text
    split: train[:500]
warmup_steps: 20
shuffle_merged_datasets: true
sample_packing: true
pad_to_sequence_len: true

# Batching
num_epochs: 2
gradient_accumulation_steps: 1
micro_batch_size: 1
eval_batch_size: 1

# Evaluation
val_set_size: 100
evals_per_epoch: 10
eval_table_size:
eval_max_new_tokens: 256
eval_sample_packing: false

save_safetensors: true

# WandB
wandb_project: Intern-Rp-Test
#wandb_entity:

gradient_checkpointing: 'unsloth'
gradient_checkpointing_kwargs:
  use_reentrant: false

unsloth_cross_entropy_loss: true
#unsloth_lora_mlp: true
#unsloth_lora_qkv: true
#unsloth_lora_o: true

# LoRA
adapter: qlora
lora_r: 32
lora_alpha: 64
lora_dropout: 0.25
lora_target_linear: true
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj
lora_modules_to_save:
#peft_use_rslora: true
#loraplus_lr_ratio: 8

# Optimizer
optimizer: paged_ademamix_8bit
lr_scheduler: cosine
learning_rate: 3e-5
cosine_min_lr_ratio: 0.1
weight_decay: 0.01
max_grad_norm: 1.0

# Misc
train_on_inputs: false
group_by_length: false
early_stopping_patience:
local_rank:
logging_steps: 1
xformers_attention:
#debug:
deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json # previously blank
fsdp:
fsdp_config:

plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_layer_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: true

gc_steps: 10
seed: 69
```

</details><br>

# intern-rp-lora

This model is a fine-tuned version of [internlm/internlm3-8b-instruct](https://huggingface.co/internlm/internlm3-8b-instruct) on the ToastyPigeon/some-rp, the BeaverAI/cedo-unalignment, the BeaverAI/foundRP, the PocketDoc/Dans-Prosemaxx-Gutenberg, the ToastyPigeon/SpringDragon-Instruct, the allenai/tulu-3-sft-personas-instruction-following and the allura-org/fujin-cleaned-stage-2 datasets.
It achieves the following results on the evaluation set:
- Loss: 1.7197

## 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: 3e-05
- train_batch_size: 1
- eval_batch_size: 1
- seed: 69
- distributed_type: multi-GPU
- num_devices: 4
- total_train_batch_size: 4
- total_eval_batch_size: 4
- optimizer: Use OptimizerNames.PAGED_ADEMAMIX_8BIT and the args are:
No additional optimizer arguments
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 20
- num_epochs: 2

### Training results

| Training Loss | Epoch  | Step | Validation Loss |
|:-------------:|:------:|:----:|:---------------:|
| 2.2794        | 0.0013 | 1    | 1.8317          |
| 1.6416        | 0.1    | 75   | 1.7826          |
| 2.3547        | 0.2    | 150  | 1.7643          |
| 1.9114        | 0.3    | 225  | 1.7546          |
| 2.0004        | 0.4    | 300  | 1.7474          |
| 2.2052        | 0.5    | 375  | 1.7428          |
| 1.9314        | 0.6    | 450  | 1.7377          |
| 2.202         | 0.7    | 525  | 1.7350          |
| 2.2453        | 0.8    | 600  | 1.7303          |
| 1.8392        | 0.9    | 675  | 1.7283          |
| 1.7018        | 1.0    | 750  | 1.7271          |
| 1.9736        | 1.0987 | 825  | 1.7264          |
| 2.0917        | 1.1987 | 900  | 1.7245          |
| 1.5679        | 1.2987 | 975  | 1.7239          |
| 2.0799        | 1.3987 | 1050 | 1.7225          |
| 1.8398        | 1.4987 | 1125 | 1.7220          |
| 1.9806        | 1.5987 | 1200 | 1.7211          |
| 1.7334        | 1.6987 | 1275 | 1.7209          |
| 2.1457        | 1.7987 | 1350 | 1.7205          |
| 1.7804        | 1.8987 | 1425 | 1.7202          |
| 2.1652        | 1.9987 | 1500 | 1.7197          |


### Framework versions

- PEFT 0.14.0
- Transformers 4.47.1
- Pytorch 2.5.1+cu124
- Datasets 3.2.0
- Tokenizers 0.21.0