README.md

---
library_name: transformers
license: llama3.2
base_model: meta-llama/Llama-3.2-3B-Instruct
tags:
- generated_from_trainer
model-index:
- name: prm-llama3.2-ToM-final
  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. -->

**Paper:** [[EMNLP'25] DEL-ToM: Inference-Time Scaling for Theory-of-Mind Reasoning via Dynamic Epistemic Logic](https://huggingface.co/papers/2505.17348)  
**Code:** [GitHub - joel-wu/DEL-ToM](https://github.com/joel-wu/DEL-ToM)

This model is part of the DEL-ToM project, which introduces a Dynamic Epistemic Logic-based framework for modeling and evaluating theory-of-mind reasoning in large language models. 

[<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: `0.4.1`
```yaml
base_model: meta-llama/Llama-3.2-3B-Instruct
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: /home/ubuntu/LLM-inference/yuheng-project/tts/ToM_PRM_final.jsonl
    conversation: llama3
    type: sharegpt
    split: "train"
    train_on_split: "train"

warmup_ratio: 0.05
val_set_size: 0.0
output_dir: ./prm-llama3.2-ToM-final
#wandb_project: preference-models
#wandb_entity: domain-generalization
wandb_watch:
wandb_name: "llama-3.2-3b-bs32_lr2e-6_prm"
wandb_log_model:

train_on_inputs: false

save_safetensors: true
#noisy_embedding_alpha: 10.0 # default for sharegpt type
dataset_prepared_path: ~/data/preference-models/last_run_prepared


dataset_processes: 48
#torch_compile: true
sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

trust_remote_code: True
adapter:
lora_model_dir:
#lora_r: 32
#lora_alpha: 16
#lora_dropout: 0.05
#lora_target_linear: true
#lora_fan_in_fan_out:




gradient_checkpointing: True

#warmup_ratio: 0.1
gradient_accumulation_steps: 4
micro_batch_size: 1
num_epochs: 1
#max_steps: 10
#optimizer: adamw_torch_fused
optimizer: paged_adamw_32bit
#lr_scheduler: constant_with_warmup
lr_scheduler: cosine
learning_rate: 2.0e-6

weight_decay: 0.0
max_grad_norm: 1.0


group_by_length: false
bf16: auto
fp16: false
tf32: true

early_stopping_patience:
local_rank:
logging_steps: 2
xformers_attention:
flash_attention: true


eval_steps:
eval_table_size:
eval_table_max_new_tokens:
save_steps: 100
save_strategy: "steps"
save_total_limit: 4
#save_safetensors: false
debug:


ddp: #true
deepspeed: #deepspeed/zero1.json # multi-gpu only

fsdp:
fsdp_config:
special_tokens:
        pad_token: <|end_of_text|>

```

</details><br>

# prm-llama3.2-ToM-final

This model is a fine-tuned version of [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) on the None dataset.

## 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: 2e-06
- train_batch_size: 1
- eval_batch_size: 1
- seed: 42
- distributed_type: multi-GPU
- gradient_accumulation_steps: 4
- total_train_batch_size: 4
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 80
- num_epochs: 1

### Training results



### Framework versions

- Transformers 4.45.2
- Pytorch 2.7.0.dev20250310+cu126
- Datasets 2.20.0
- Tokenizers 0.20.3