Spaces:
Running
Running
Commit History
-
README: add model card highlights section and metadata snapshot. Running
Made-with: Cursor279d788 -
README: add Hugging Face model link to Quick Links.
Made-with: Cursor3d1b780 -
Use .git GitHub URL in HTML and Gradio links.
Made-with: Cursor72d394b -
Gradio: update W&B line to 'example run' dashboard.
README: major refresh with benchmark snapshot, evidence/proof tables, artifact links, architecture, and clearer quick links. Made-with: Cursor6674354 -
Demo copy: drop judges wording; workflow map at a glance for readers.
Made-with: Cursor270cdf0 -
Expand problem narrative and Engineering Notes: time-on-SQL, Spider vs prod.
HTML /demo: stat callout lede, longer blog with subheads, lists, footnote on ranges. Gradio blog: matching cost/benchmark sections and how-to-read guide. Made-with: Cursor40caa50 -
Add GitHub repo link; drop Karpathy, blog, slides, demo placeholders.
Links to https://github.com/mdayan8/sql-debug-env in demo header, footer, engineering notes, and Gradio first-context + references. Made-with: Cursord00292e -
HTML demo: fix diagram crop; click-to-zoom lightbox with pan and +/- zoom.
Diagram uses scrollable wrap without max-height crop; all figures zoomable with full-size viewer, Ctrl+wheel zoom, Esc/backdrop close. Made-with: Cursore6d1a8f -
HTML /demo: add Benchmark visuals table and three chart figures.
Nav link to #benchmark-visuals; metric snapshot table; chart-performance-leap, chart-comparison-shift, chart-spider-benchmark with captions. Made-with: Cursorf5c939b -
Gradio: fix invisible header and body text on HF dark embed.
- Single HTML hero so styling is not split across Column DOM. - Force light prose/heading colors; fix .sde-section-title (was dark ink on dark bg). - Container default text color for theme gaps; link and caption tweaks. Made-with: Cursor293388d -
Space home: redirect / to HTML /demo; Gradio at /gradio; fix Gradio hero.
- Root loads the demo page; demo nav and banner link to /gradio/. - Rebuild hero as Column + Markdown so title and pills render under Gradio 5 themes. - CSS: force light-on-dark text inside .sde-hero-wrap for prose overrides. Made-with: Cursorf4ae3f3 -
Fix Gradio 5 mount error by moving theme/css to Blocks.
mount_gradio_app in Gradio 5 does not accept theme/css kwargs, causing startup failure on Space. Made-with: Cursor3cc7ab5 -
Fix Space pip resolve: fastapi>=0.115.2, python-multipart>=0.0.18, pin gradio 5.50.0.
Resolves conflicts between pinned fastapi 0.115.0 / multipart 0.0.9 and Gradio 5.x deps. Made-with: Cursord9b6c59 -
Pin Gradio >=5.7.1 for huggingface_hub 1.x (fixes HfFolder ImportError on Space).
Made-with: Cursorf7153ad -
Point /demo and Gradio at diagram-end-to-end-workflow.png (asset on Hub via Xet).
Made-with: Cursor4c3e70f -
Add workflow diagram via Hub API (Xet)
Verified signature096e16d -
Restore optimized training diagnostics and reward curve images
Made-with: Cursor35a3454 -
Ship polished Space UI with Gradio dashboard and evidence-rich demo.
Add the upgraded Gradio experience, structured blog/reporting sections, and curated artifact images while keeping OpenEnv API endpoints intact. Made-with: Cursor029f9cf -
Upload artifacts/runs/20260426-064318-sample-rewards-32eval/sample_rewards_final.json with huggingface_hub
Verified signature4724001 -
Upload folder using huggingface_hub
Verified signatured2042c8 -
Add training artifacts 20260426-050707
Verified signature05bd818 -
Upload folder using huggingface_hub
Verified signature0730521 -
Upload folder using huggingface_hub
Verified signature4ca17fb -
Add training artifacts 20260426-032246
Verified signatured153683
-
Add training artifacts 20260425-202751
Verified signature471dda7 -
Avoid whoami rate-limit during job submission
Use HF_JOB_NAMESPACE default to md896 and skip whoami lookup so jobs submit reliably when whoami-v2 is rate-limited.a1e637f -
Harden HF job token wiring and persist full training outputs
Pass a real HF token secret into Jobs, push full trained weights/tokenizer, and upload run artifacts to the Space repo. Add explicit hard-task base-vs-trained reward evaluation so performance deltas are saved with graphs and metrics.9552aaf -
Fix TRL 0.18 compatibility: remove unsupported generation_kwargs; set safety flags on model.generation_config.
Made-with: Cursor6083a40 -
Harden GRPO generation stability on CUDA: bf16 + eager attention + invalid-logit guards.
Made-with: Cursor948530a -
Fix GRPO batch/generation mismatch: auto-adjust num_generations; set launcher default to 2.
Made-with: Cursoraf54ccd -
Simplify HF training stack: remove unsloth/vllm path, use plain transformers AutoModel + single OpenEnv reward.
Made-with: Cursore5262a1 -
Fix Unsloth startup: avoid pre-importing trl/transformers; mock vllm as real package modules.
Made-with: Cursord21de11 -
Fix HF job startup: import unsloth first and shim vllm package metadata check.
Made-with: Cursor1fdba13 -
Fix HF Job bootstrap: transformers>=4.51 for trl 0.18, datasets<4; simplify to colab-style OpenEnv SQL reward.
Made-with: Cursoree30276 -
Fix HF Jobs bootstrap (pin transformers/trl, drop torchao stack); add reward and trainer JSONL logging; stabilize launch_job.
Made-with: Cursorceee0e3 -
changes in ultimate sota
ac3911c -
Fix: Mock vllm and llm_blender to stabilize GRPOTrainer in HF Jobs environment
bc20ef9 -
Downgrade TRL to 0.22.2 to natively bypass experimental vllm dependencies
2eb9add -
Fix vllm error cleanly by creating fake python module structure
b2ce6c6 -
Add vllm to dependencies to fix TRL's hard import requirement
711ae38 -
Remove vllm mock to fix importlib find_spec crash in TRL 0.23
97cddc4 -
Mock vllm to bypass TRL missing module error
1bc1daa -
Downgrade TRL to <0.24.0 to fix missing dependency chain
397face -
Add TRANSFORMERS_CACHE mock to fix TRL/llm_blender crash
16dd181 -
Fix llm-blender ModuleNotFoundError
1cd9ac8 -
Ensure TRL GRPO imports by installing mergekit
TRL's GRPO trainer path imports mergekit utilities unconditionally in recent releases; without mergekit present the job fails before training starts.\n\nInstall mergekit during bootstrap so GRPOConfig/GRPOTrainer imports succeed in HF Jobs containers. Constraint: HF Jobs run with ephemeral, system-managed Python (PEP-668)\nRejected: Pin older TRL | risk of drifting from Unsloth-supported TRL range\nConfidence: medium\nScope-risk: narrow\nDirective: If TRL relaxes this dependency again, mergekit can be removed to reduce install time\nTested: python -m py_compile ultimate_sota_training.py\nNot-tested: HF Jobs end-to-end run6d0b5c3 -
Prevent torchvision import crashes in HF Jobs
HF base images can include torchvision, and pip uninstall may be blocked by PEP-668 unless break-system-packages is enabled. When Unsloth's dependency resolution downgrades torch, any remaining torchvision can crash transformers/trl imports (torchvision::nms).\n\nSet TRANSFORMERS_NO_TORCHVISION early, opt into PIP_BREAK_SYSTEM_PACKAGES for all pip ops, and uninstall torchvision/torchaudio both before and after Unsloth install so text-only training stays stable. Constraint: HF Jobs run on externally-managed system Python (PEP-668)\nRejected: Pin torch/torchvision versions | pip resolver still drags heavy CUDA wheels and increases cost\nConfidence: high\nScope-risk: narrow\nDirective: Keep torchvision out of text-only training images unless you also pin torch+torchvision compat\nTested: python -m py_compile ultimate_sota_training.py\nNot-tested: HF Jobs end-to-end run8b3c03a -
Make OpenEnv training+API judge-proof
Hugging Face Jobs runs were failing due to torch/torchvision mismatches triggered by dependency resolution. The GRPO training script now avoids optional vision deps for text-only runs and emits real artifacts (log history + reward curve + sampled before/after execution reward) instead of illustrative charts. Also hardens the reviewer flow and aligns the public contract: adds a state->observation builder for reviewer rejections, keeps reviewer rewards inside strict (0,1), updates the manifest + README for the finance task, and adds socketless API integration tests via FastAPI TestClient. Restores a root-level baseline inference runner as documented. Constraint: HF Jobs images may ship torch/torchvision stacks that become incompatible after pip resolution Constraint: Judges need rerunnable training evidence (plots/logs) sourced from real runs Rejected: Force-pin torch/torchvision via pip | large downloads and brittle across images Confidence: high Scope-risk: moderate Reversibility: cl…d061422 -
Add --break-system-packages for Ubuntu 24.04 (PEP-668)
830c039 -
Fix TRL and Torchao dependency conflicts
d118f9f