- Stop llama-server subprocess on model unload instead of relying on GC
- Fix tool_calls[].index being string instead of int in API responses
- Omit tool_calls key from API response when empty per OpenAI spec
- Prevent division by zero when micro_batch_size > batch_size in training
- Copy sampler_priority list before mutating in ExLlamaV3
- Normalize presence/frequency_penalty names for ExLlamaV3 sampler sorting
- Restore original chat_template after training instead of leaving it mutated
The tool call response parser only handled JSON-based formats, causing
tool_calls to always be empty for models that use non-JSON formats.
Add parsers for three additional tool call formats:
- Qwen3.5: <tool_call><function=name><parameter=key>value</parameter>
- Mistral/Devstral: functionName{"arg": "value"}
- GPT-OSS: <|channel|>commentary to=functions.name<|message|>{...}
Also fix multi-turn tool conversations crashing with Jinja2
UndefinedError on tool_call_id by preserving tool_calls and
tool_call_id metadata through the chat history conversion.
If --user-data-dir is not set, auto-detect: use ../user_data when
./user_data doesn't exist, making it easy to share user data across
portable builds by placing it one folder up.
Instead of two separate paths (format files vs Chat Template), all
instruction training now uses apply_chat_template() with assistant-only
label masking. Users pick a Jinja2 template from the dropdown or use the
model's built-in chat template — both work identically.
- Remove pre-padding from tokenize() and tokenize_conversation()
- Collate function now right-pads each batch to the longest sequence
- Set tokenizer padding_side to "right" (standard for training)
- Remove dead natural_keys import
- Reduces wasted compute on batches with short sequences
- Aligns with axolotl/unsloth approach
- Replace "Raw text file" tab with "Text Dataset" tab using JSONL format with "text" key per row
- Re-add stride overlap for chunking (configurable Stride Length slider, 0-2048 tokens)
- Pad remainder chunks instead of dropping them
- Remove hard_cut_string, min_chars, raw_text_file parameters
- Remove .txt file and directory loading support
- Support multi-turn conversations (OpenAI messages + ShareGPT formats)
- Automatic assistant-only label masking via incremental tokenization
- Use tokenizer.apply_chat_template() for proper special token handling
- Add "Chat Template" option to the Data Format dropdown
- Also accept instruction/output datasets (auto-converted to messages)
- Validate chat template availability and dataset format upfront
- Fix after_tokens[-1] IndexError when train_only_after is at end of prompt
- Update docs
- Use peft's "all-linear" for target modules instead of the old
model_to_lora_modules mapping (only knew ~39 model types)
- Add "Target all linear layers" checkbox, on by default
- Fix labels in tokenize() — were [1]s instead of actual token IDs
- Replace DataCollatorForLanguageModeling with custom collate_fn
- Raw text: concatenate-and-split instead of overlapping chunks
- Adapter backup/loading: check safetensors before bin
- Fix report_to=None crash on transformers 5.x
- Fix no_cuda deprecation for transformers 5.x (use use_cpu)
- Move torch.compile before Trainer init
- Add remove_unused_columns=False (torch.compile breaks column detection)
- Guard against no target modules selected
- Set tracked.did_save so we don't always save twice
- pad_token_id: fall back to eos_token_id instead of hardcoding 0
- Drop MODEL_CLASSES, split_chunks, cut_chunk_for_newline
- Update docs
