Spaces:
Sleeping
Sleeping
Commit History
-
fix: set per_device_eval_batch_size=1 to prevent OOM during evaluation loop
26efb80 -
perf: increase batch size to 2 and enable group_by_length to massively speed up training
c38b6f2 -
fix: change device_map='auto' to local_process_index to prevent cross-GPU loss calculation error in transformers 4.46
b602c40 -
fix: set remove_unused_columns=True so Trainer drops 'messages' list before data collation
1bbece6 -
fix: change max_length to max_seq_length in SFTConfig for trl 0.12.1 compatibility
930fd33 -
refactor: address Phase 2 review findings
- #3 add length filter: drop 82 train examples >2560 tok (verified) so their SQL labels aren't truncated; dev unaffected (0 dropped). MAX_LEN constant (DRY vs SFTConfig). - #4 DRY: load_dataset now uses train_file/val_file vars instead of hardcoded paths. - #1 document + guard the peft _torch_supports_distributed test workaround (hasattr-gated). - #5 fix stale comments (grad_accum 4->8; 'all linear layers' -> fused-projection rationale). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>e03c947 -
fix: LoRA adapts all 7 projections via Phi-3 fused names + guard assertion; r=16, Spider data, max_length=2560
5f0c197 -
feat: Spider data pipeline pinned to minktn/spider-data (per-example CREATE TABLE schema); drop find_spider.py
2ff9360
-
Complete CPU GGUF serving + docs + minimal UI redesign
Serving: - Serve fine-tuned Phi-3 as a 4-bit Q4_K_M GGUF via llama.cpp (source-built llama-cpp-python 0.3.19); cap n_threads to the container's vCPU count. - Dockerfile builds llama-cpp-python from source against glibc (prebuilt wheels are musl-only for phi3-capable versions) with single-threaded compile to avoid build OOM. Frontend: - Redesign the web UI as a flat, minimal layout (no gradients/glass/glow), with a project stats panel, schema row counts, and an on-page FAQ. - Remove the unreliable base-model comparison; single fine-tuned result with a wall-clock timer and an accurate CPU banner. Docs + eval: - Add fine-tuning + quantization + deployment blog and HF model cards (docs/). - Add the execution-accuracy eval harness (scripts/eval_execution_accuracy.py). - Update repo README to reflect GGUF/CPU serving. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>1d8be28
-
Revert "Migrate CPU serving to llama.cpp + Q5_K_M GGUF (was bf16 transformers)"
This reverts commit 564ad28ca5effbb5f2390da477c1915c03fbc8ad.d031aeb -
Revert "Make torch/transformers/peft lazy imports so CPU Space boots without them"
This reverts commit eac26c792828354a0af21d78d763b0320eb0d2b2.2f03b0e -
Make torch/transformers/peft lazy imports so CPU Space boots without them
eac26c7 -
Migrate CPU serving to llama.cpp + Q5_K_M GGUF (was bf16 transformers)
564ad28 -
Fix bugs
52e3716 -
Fix bugs
19cfcb8 -
Update CPU warning banner with accurate load and query times
e4ff54d -
Fix inference device bug, auto-load model, remove fine-tuning console, UI cleanup
f5ce94d -
initial deployment
55159b1