P1 #1: ``_gpu_workload_busy_for_helper`` in
``utils/datasets/llm_assist.py`` now also gates on the GGUF chat
backend (llama-server) AND the safetensors chat backend. Round 23
extended it to training + export but missed Chat, so a helper /
advisor GGUF could still race a loaded chat model for VRAM.
Both checks fail closed when status is unverifiable.
P1 #2 / #3 / #4 / #5: re-ordered the route-level GPU-handoff
unloads so the diffusion release runs BEFORE the chat releases.
A wedged diffusion unload used to fire AFTER chat was already
gone, so the user lost both on a single failure. Drop chat last
so an earlier failure preserves it. Applied to
``/training/start`` (training.py), ``/export/load`` (export.py),
``/chat/load`` GGUF branch and ``/chat/load`` safetensors branch
(routes/inference.py).
P1 #7 + P2 #13: ``/delete-finetuned`` body now hardens
``model_path`` and ``gguf_variant`` via the shared
``_validate_logged_identifier`` helper, so control characters
and URL-form HF tokens can no longer log-line-smuggle.
P1 #8 + #10: ``/delete-cached`` body hardens ``repo_id`` and
``variant`` the same way.
P1 #9: ``/download-progress`` ``repo_id`` query parameter is
also hardened; the value flows into log lines deep inside
``_get_repo_size_cached`` on lookup failure.
P1 #11: ``CheckFormatRequest.dataset_name`` and
``AiAssistMappingRequest.{dataset_name, model_name}`` in
``models/datasets.py`` now apply the same control-char +
embedded-HF-token validators, matching every other public
request-body model.
All 115 diffusion + training-validation + cached_gguf + export
+ inference model-validation tests pass locally.
(P1 #6 native-path-lease enforcement for diffusion local paths
and P1 #12 React Compiler frontend lint deferred -- both need
focused design / frontend touchups separate from this batch.)
The advisor now only assigns columns to user/assistant roles and
generates a system prompt. Templates (user_template, assistant_template)
are removed entirely — the LLM was frequently putting all columns in
user or copying actual data values into templates.
Column values are now used directly as message content, grouped and
concatenated by role. This is simpler, more robust, and prevents the
class of bugs where the advisor generates bad template content.
Non-conversational HF datasets (e.g. stanfordnlp/snli) were naively mapped
column→role, producing poor training results. The AI Assist button now runs
a 3-pass advisor using Qwen 7B that:
1. Fetches the HF dataset card/README to understand the dataset purpose
2. Classifies the dataset type and determines if conversion is needed
3. Generates a system prompt, user/assistant templates with {column}
placeholders, and label mappings (e.g. 0→entailment)
4. Validates the conversion quality (score ≥7/10 required)
Architecture: advisor metadata flows as __-prefixed keys in
custom_format_mapping (e.g. __system_prompt, __user_template,
__assistant_template, __label_mapping). The existing _apply_user_mapping()
detects these keys and routes to template-based conversation construction.
No __ keys = existing simple mode (backwards compatible).
Backend: upgraded llm_assist.py (7B default, multi-pass advisor,
HF card fetching), extended API models, added _apply_template_mapping()
to dataset_utils.py.
Frontend: extended store with advisor state fields, wired AI Assist
to store templates/system prompt, inject __ metadata in training request,
show advisor notification banner in mapping card.
Move LLM-assisted column mapping from silent /check-format automation
to an explicit "AI Assist" button in the dataset mapping dialog. This
makes the feature transparent and user-controlled.
- Remove llm_classify_columns() from check_dataset_format() (heuristic-only)
- Remove auto-save suggested_mapping from use-training-actions.ts
- Add POST /api/datasets/ai-assist-mapping endpoint (receives preview
samples from frontend, no dataset re-loading needed)
- Add AiAssistMappingRequest/Response models
- Add aiAssistMapping() frontend API function
- Add Sparkles AI Assist button to DatasetMappingCard with loading state
- Wire up handleAiAssist handler in dataset-preview-dialog.tsx
- Add 200-sample parallel probe using ThreadPoolExecutor + safe_num_proc
to estimate download speed and failure rate before full conversion
- Abort with clear error if >=30% of probe images fail to download
- Show estimated download time in the training overlay modal
- Parallel batch conversion for URL-based datasets (vs sequential for local)
- Add warning field to /check-format response for URL-based image datasets
- Display URL warning in dataset preview dialog (amber banner)
- Thread progress_callback from trainer through format_and_template_dataset
to convert_to_vlm_format for real-time status updates
