See axolotl config

axolotl version: 0.7.0

base_model: meta-llama/Llama-3.2-3B-Instruct
hub_model_id: morsmordre/m-3b-v1-iteration-00-sf-xlam-06

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: minpeter/xlam-function-calling-60k-hermes
    data_files:
      - result.parquet
    type: chat_template
    chat_template: llama3
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    shards: 30
  - path: minpeter/xlam-irrelevance-7.5k-qwen2.5-72b-distill-hermes
    data_files:
      - result.parquet
    type: chat_template
    chat_template: llama3
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    shards: 6
  - path: minpeter/bfcl-v1-non-live-ast-hermes
    data_files:
      - result.parquet
    type: chat_template
    chat_template: llama3
    field_messages: conversations
    message_field_role: from
    message_field_content: value

chat_template: llama3

dataset_prepared_path: last_run_prepared

output_dir: ./output

adapter: lora
lora_model_dir:

sequence_len: 4096
pad_to_sequence_len: true
sample_packing: true

val_set_size: 0.05
eval_sample_packing: true
evals_per_epoch: 3

lora_r: 8
lora_alpha: 16
lora_dropout: 0.05
lora_fan_in_fan_out:
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj

wandb_project: "axolotl"
wandb_entity: "kasfiekfs-e"
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 2
micro_batch_size: 2
num_epochs: 1
optimizer: adamw_8bit
lr_scheduler: cosine
learning_rate: 0.0002

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3

warmup_steps: 10
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:

special_tokens:
  pad_token: "<|finetune_right_pad_id|>"

m-3b-v1-iteration-00-sf-xlam-06

This model is a fine-tuned version of meta-llama/Llama-3.2-3B-Instruct on the minpeter/xlam-function-calling-60k-hermes, the minpeter/xlam-irrelevance-7.5k-qwen2.5-72b-distill-hermes and the minpeter/bfcl-v1-non-live-ast-hermes datasets. It achieves the following results on the evaluation set:

• Loss: 0.1362

Model description

Test Type (bfcl)	Base Model Accuracy	This Adapter Accuracy	Improvement
irrelevance	72.08	74.17	+2.09
parallel_multiple	10.00	90.00	+80.00
parallel	11.50	92.00	+80.50
simple	24.75	95.00	+70.25
multiple	20.00	93.50	+73.50

Inference

vllm serve meta-llama/Llama-3.2-3B-Instruct \
--enable-auto-tool-choice --tool-call-parser llama_hermes --tool-parser-plugin github.com/minpeter/hermes-llama-parse/lh_tool_parser.py  \
--port 4000 --enable-lora --lora-modules tool='morsmordre/m-3b-v1-iteration-00-sf-xlam-06'

The hermes parser of the existing vllm expects the text and to exist as a single token in the vocab, so a modified "llama_hermes" parser is required. As of v0.7.3, you can inject the plugin using the --tool-parser-plugin flag.

Model Evaluation Results

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: 0.0002
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 4
• optimizer: Use OptimizerNames.ADAMW_8BIT with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 1.0

Training results

Training Loss	Epoch	Step	Validation Loss
0.6162	0.0059	1	0.4366
0.2323	0.3343	57	0.1475
0.1065	0.6686	114	0.1362

Framework versions

• PEFT 0.14.0
• Transformers 4.48.3
• Pytorch 2.5.1+cu124
• Datasets 3.2.0
• Tokenizers 0.21.0