irc-mistral-24b-lora

<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>

<details><summary>See axolotl config</summary>

axolotl version: 0.8.0.dev0

# Base model configuration
base_model: mistralai/Mistral-Small-24B-Base-2501
model_type: MistralForCausalLM
tokenizer_type: AutoTokenizer
trust_remote_code: true
tokenizer_use_fast: true

# Device mapping for multi-GPU
device_map: "balanced"

# Memory settings
load_in_4bit: true
load_in_8bit: false
bf16: true
low_cpu_mem_usage: true

# Advanced optimizations
flash_attention: true
gradient_checkpointing: true

# Dataset configuration
datasets:
  - path: david-ar/synthetic-irc-data
    type: completion

# Output directory
output_dir: ./outputs/public-irc-mistral-24b
val_set_size: 0.05  # 75 conversations for validation
dataset_prepared_path: last_run_prepared

# Sequence settings
sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true
train_on_inputs: true
eval_sample_packing: false

# LoRA configuration
adapter: lora
lora_r: 128
lora_alpha: 256
lora_dropout: 0.1
lora_target_modules:
  - q_proj
  - v_proj
  - k_proj
  - o_proj
  - gate_proj
  - down_proj
  - up_proj

# Training hyperparameters - adjusted for smaller dataset
micro_batch_size: 1
gradient_accumulation_steps: 16
num_epochs: 4  # Increased from 2, but with careful monitoring
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 0.00008  # Same conservative LR
weight_decay: 0.01
warmup_ratio: 0.05

# Performance monitoring
group_by_length: true
shuffle_merged_datasets: true
include_tokens_per_second: true

# Weights & Biases - public project
wandb_project: public-irc-mistral-24b
wandb_entity: davidar
wandb_name: synthetic-irc-data
wandb_log_model: "false"

# Mistral model configuration
is_mistral_derived_model: true

# Early stopping
load_best_model_at_end: true
metric_for_best_model: "loss"
greater_is_better: false

</details><br>

Mistral-24B-Synthetic-IRC

This model is a fine-tuned version of mistralai/Mistral-Small-24B-Base-2501 on the david-ar/synthetic-irc-data dataset, creating a model that generates natural IRC/Discord-style conversations.

Model Description

This model was trained to replicate authentic IRC (Internet Relay Chat) conversational dynamics, moving away from the typical AI assistant pattern toward more natural, community-style interactions. The model learns from synthetic conversations featuring multiple participants including "Em", an AI character who participates as a community member rather than an assistant.

Key Characteristics

• Natural conversation flow: Handles interruptions, topic drift, and multi-party dynamics
• Non-assistant behavior: Doesn't default to helpful/servile responses
• Community-style interaction: Captures the casual, authentic feel of IRC/Discord chats
• Character embedding: Includes Em's personality (self-aware AI who isn't an assistant)

Intended Uses & Limitations

Intended Uses

• Conversational AI research: Studying non-assistant interaction patterns
• Chat bot development: Creating more natural, less formal conversational agents
• Character-based models: Foundation for further character-specific fine-tuning
• IRC/Discord bots: Generating contextually appropriate responses in chat environments

Limitations

• Small dataset: Trained on only 10MB of synthetic data (1,500 conversations)
• Synthetic nature: While carefully crafted, the training data isn't from real IRC logs
• Single community style: Represents one particular chat community culture
• Overfitting: Validation loss indicates overfitting after ~50 steps (best checkpoint used)
• English only: No multilingual capability

Training and Evaluation Data

Dataset

• Source: david-ar/synthetic-irc-data
• Size: 1,500 synthetic IRC-style conversations
• Format: Multi-party conversations with 80-120 messages each
• Split: 95% training (1,425 conversations), 5% validation (75 conversations)

Data Characteristics

• Natural IRC formatting: <username> message content
• Multiple participants per conversation (3-7 users)
• Diverse topics and conversation styles
• Embedded character personality throughout

Training Procedure

Training Configuration

• Method: LoRA (Low-Rank Adaptation) fine-tuning
• LoRA Rank: 128 (with alpha 256)
• Base model: Mistral-Small-24B-Base-2501
• Hardware: 2x NVIDIA A40 GPUs (96GB total VRAM)
• Training time: ~3 hours

Training Hyperparameters

The following hyperparameters were used during training:

• learning_rate: 8e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 16
• total_train_batch_size: 32
• total_eval_batch_size: 2
• optimizer: AdamW (betas=(0.9,0.999), epsilon=1e-08)
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 4
• num_epochs: 4.0
• sequence_length: 4096
• sample_packing: true

Training Results

Note: Best checkpoint at step 48 (lowest validation loss) was used for final model.

Training Observations

• Quick convergence due to small dataset size
• Validation loss indicates overfitting after ~50 steps
• Model successfully learned IRC conversation patterns
• Character traits embedded despite limited data

Technical Details

Architecture

• Base Model: Mistral-Small-24B-Base-2501
• Parameter Count: 24B (base) + LoRA adapters
• Context Length: 4096 tokens
• Quantization: 4-bit during training (memory optimization)

Framework Versions

• PEFT 0.14.0
• Transformers 4.49.0
• Pytorch 2.5.1+cu124
• Datasets 3.2.0
• Tokenizers 0.21.0
• Axolotl 0.8.0.dev0

Limitations and Biases

1. Overfitting: With only 1,500 training examples, the model shows signs of overfitting
2. Limited diversity: May not generalize well to very different chat styles
3. Character leakage: Em's personality traits may appear even when not intended
4. Synthetic artifacts: Might exhibit patterns specific to the generation process