CogStateIR-Supra50M-v0.0.2-small_dataset

Supra 50 M Instruct (second time)

The model is available here.

Training setup

• Base model : https://huggingface.co/SupraLabs/Supra-50M-Instruct
• Parameters : 51.8M
• Dataset : 135 examples
• Epochs : 100
• Learning rate : 0.0001
• Hardware : Apple M2 Pro on GPU
• Duration : ~20 minutes (thanks to Metal)

Results

Loss after training : 0.397188. Validation loss after training: 1.135684.

Prediction - training example

Model's raw output :

"emotion|anger|increases_a_little\nemotion|hurt|increases\nrelationship|closeness|increases\nrelationship|trust|increases_a_little\nbelief|user_is_dismissive|increases"

Model's prediction :

state_changes:
  emotion:
    anger: increases_a_little
    hurt: increases
  relationship:
    closeness: increases
    trust: increases_a_little
  belief:
    user_is_dismissive: increases
  memory: null
  reflection: null

Output of the dataset :

state_changes:
  relationship:
    trust: increases_a_little
  emotion:
    anger: decreases
  belief:
    people_can_change: increases_a_little
  memory: reinforce_previous_conflict
  reflection: required

The model didn't really integrated the difference between relationship and emotion, didn't really integrated the subtleties about the character personnality, but didn't totally hallucinated : he integrated the output format, didn't repeat himself, and proposed a pretty relevant IR (not totally absurd). The result is quite the same as for the first model, but a little better (more content).

Prediction - held-out example

Model's raw output :

RAW OUTPUT:
"emotion|anger|increases\nemotion|fear|increases\nemotion|fear|increases\nrelationship|fear|increases\nrelationship|fear|increases\nrelationship|fear|increases\nrelationship|trust|increases\nbelief|user_cant_take|increases\nreflection|required"

Model's output :

state_changes:
  emotion:
    anger: increases
    fear: increases
  relationship:
    fear: increases
    trust: increases
  belief:
    user_cant_take: increases
  memory: null
  reflection: required

Output of the test dataset :

state_changes:
  relationship:
    respect: increases_a_little
    disappointment: decreases
  emotion:
    irritation: decreases
    pride: increases_a_little
  belief:
    effort_matters: increases
  memory: reinforce_previous_conflict
  reflection: required

As you can see, the modeld didn't hallucinated this time : he didn't repeated himself, gave a substancial output compared to the one of the first model. We still have the problem of the pertinence of the changes. I think this can be because of :

• the small size of the model
• the small size of the dataset

Next steps

Next things to do are :

• train a larger model (~360M parameters) with a larger dataset (~500 examples) and ~100 epochs (to avoid overtraining and memorization)
• train a LoRA instead of finetuning all the model.
• train with an even largers datasets : 250, 500, 1500, 2500 examples.
• evaluate the model's predictions over many tests
• integrate the IR in the engine