drop length column for issues with eval without packing (#1711)

download model weights on preprocess step (#1693)

verbose failure message (#1694)

fix for when sample_packing and eval_sample_packing are different (#1695)

add back packing efficiency estimate so epochs and multi-gpu works properly (#1697)

ensure explicit eval_sample_packing to avoid mismatch issues (#1692)

Phi-3 conversation format, example training script and perplexity metric (#1582)

add support for rpo_alpha (#1681)

re-enable DPO for tests in modal ci (#1374)

need to add back drop_last for sampler (#1676)

cleanup the deepspeed proxy model at the end of training (#1675)

load explicit splits on datasets (#1652)

set chat_template in datasets config automatically (#1664)

use mixins for orpo and kto configs so they work with axolotl customizations (#1674)

revert multipack batch sampler changes (#1672)

handle the system role too for chat templates (#1671)

make sure the CI fails when pytest script fails (#1669)

Correct name of MixtralBlockSparseTop2MLP (L -> l) (#1667)

Fix: ensure correct handling of `val_set_size` as `float` or `int` (#1655)

Generalizing the chat_template prompt strategy (#1660) [skip ci]

Switch to parallel FFD bin packing algorithm. (#1619)

support for custom messages field in sharegpt (#1651)

enable loraplus setting for dpo trainer (#1646)

allow report_to for multiple providers (#1647)

Fix llama3 chat_template (extra <|eot_id|> on last turn) (#1635)

Add KTO support (#1640)

fixes to save on fractional save_steps (#1643)

Unsloth optims for Llama (#1609)

add save_only_model option (#1634)

Fix `total_num_steps` (#1566)

FIX: max_length and max_prompt_length was not being sent to ORPOTrainer (#1584)

make sure to save on the last step (#1615)

fix attention mask collation (#1603)

Llama3 dpo (#1610)

feat: Add LLaMA-3 instruct prompt strategies for fine-tuning (#1553)

adding llama3 fastchat conversation monkeypatch (#1539)

ignore the fsdp_config section too (#1606) [skip ci]

make sure to save the lora adapter at the end of RL/dpo training (#1573)

improve tool handling roles (#1587)

feat: exclude mamba blocks for jamba (#1578)

Pass deepspeed and fsdp as None explicitly when merging adapters to allow custom device_map (#1575)

Gradio configuration parameters (#1591)

improve save callbacks (#1592)

Pass weakref to model in the SIGINT handler to free up model post train function (#1581)

FIX: TRL trainer preprocessing step was running in one process (#1583)

ADD: warning hub model (#1301)

Add debug option for RL dataset preprocessing (#1404)

PoSE context length ext (#1567)

make sure everything stays in the same dtype when using dpo + FSDP (#1559)

Add support for Gemma chat template (#1530)