Spaces:
Sleeping
Sleeping
Commit History
-
Add gated-repos OAuth scope for GAIA dataset attachment access
Verified signature3c71c18 -
Add huggingface_hub to requirements.txt
Verified signature8702001 -
Fetch GAIA attachments directly from the dataset, bypass broken proxy
Verified signatureScoring server's /files/{task_id} always 404s (confirmed bug, matching open unmerged PR on agents-course/Unit4_scoring). Fetch attachments directly from gaia-benchmark/GAIA dataset at 2023/validation/<file_name> via hf_hub_download() with the user's OAuth token instead. Verified working for png/mp3/py.d12b99d -
Switch to zai-glm-4.7: gpt-oss-120b's daily quota is exhausted
Verified signatureConfirmed via Cerebras dashboard: gpt-oss-120b hit 1.07M/1M tokens today. Each free-tier model has its own separate daily quota; switch to zai-glm-4.7 for an untouched budget.e3358fb -
Revert max_steps to 7, stop wasting steps on Wikipedia direct visits
Verified signaturemax_steps=12 blew Cerebras' daily token quota partway through the run, auto-failing the back half of questions (25% -> 10%). Reverted to 7. Also added an explicit instruction against visit_webpage on wikipedia.org URLs, which always 403s - use wikipedia_search instead.43ea6cc -
Increase max_steps 7 -> 12 to reduce premature cutoffs
Verified signature25% run showed questions hitting max_steps while still making progress, forcing garbled non-tool-call final answers. TokenPacer sleeps stayed short (6-8s), so there's TPM headroom to let runs go longer.9190f43 -
Switch from CodeAgent to ToolCallingAgent
Verified signature15% run showed gpt-oss-120b frequently breaking CodeAgent's <code> tag parser, burning steps on parse failures instead of progress. ToolCallingAgent uses native structured tool calls instead, eliminating this failure mode. PythonInterpreterTool added explicitly to retain compute capability.2a57317 -
Switch model backend from Groq to Cerebras
Verified signatureGroq's free-tier TPM (6000-12000) was too small for this agent's per-call cost (2400-9000+ tokens). Cerebras free tier offers 30,000 TPM / 1,000,000 TPD - several times more headroom. Uses CEREBRAS_API_KEY, model cerebras/gpt-oss-120b.31039a7 -
Re-add token-aware pacing for llama-3.1-8b-instant's tight TPM
Verified signature8b model's 6000 TPM cap is tighter in practice than 70b's - a single call already costs 2400-4000 tokens before any tool output. Native requests_per_minute can't account for token size. Re-add TokenPacer step_callback, drop redundant add_base_tools DuckDuckGo tool to shrink per-call overhead.974b41a -
Switch to llama-3.1-8b-instant: fresh, larger daily quota
Verified signaturellama-3.3-70b-versatile's 100K tokens/day free-tier quota was exhausted by our own testing (98090/100000 used). Groq tracks quota per model - llama-3.1-8b-instant has a separate 500K/day budget, untouched.741cd88 -
Use smolagents' native RPM/retry controls, cap steps, trim memory
Verified signatureCustom step_callback throttling couldn't see smolagents' own internal retries on rate-limit errors. Switch to LiteLLMModel's native requests_per_minute=25 and retry=False. Cap max_steps=7 and add MemoryTrimmer to stop unbounded context growth (the real driver of TPM failures). Prioritizes speed - a failed question is skipped, not retried.05bcb11 -
Make rate limiter token-aware; trim verbose tool outputs
Verified signatureGroq's 429s were a TPM (12k/min) limit, not RPM. RateLimiter now tracks actual token usage per step in a trailing 60s window; VisitWebpageTool and WikipediaSearchTool outputs trimmed to reduce per-call token volume.8be6e91 -
Switch model backend from Gemini to Groq
Verified signatureGemini free tier caps at 20 requests/day for gemini-3.6-flash - too low for a multi-step agent benchmark. Groq's llama-3.3-70b-versatile free tier allows 1000/day, 30 RPM. Uses GROQ_API_KEY.eebc9ca -
Fix rate limiter to throttle every call, not just after N
Verified signatureConfirmed via AI Studio dashboard: gemini-3.6-flash free tier caps at 5 RPM. Old limiter allowed unthrottled bursts up to 15 calls before pausing, which 429'd. New limiter enforces a minimum delay between every single call.c9fc314 -
Switch model backend from Claude to Gemini
Verified signatureUses GEMINI_API_KEY (free tier) instead of ANTHROPIC_API_KEY, model_id gemini/gemini-3.6-flash via LiteLLMModel.8b1ef20 -
Drop unused inference-api OAuth scope
Verified signature4fca1d8 -
Add litellm extra back to requirements.txt
Verified signature4a8afa4 -
Switch back to Claude via LiteLLM, add step rate limiter
Verified signatureFree HF Inference Providers OAuth access was 403ing; switch back to LiteLLMModel/ANTHROPIC_API_KEY, add a step_callback rate limiter to avoid provider throttling across a full eval run.155f4b7
-
Add inference-api OAuth scope for per-user Inference Providers calls
Verified signature222b65b -
Drop litellm extra, no longer needed
Verified signature7f22ef3 -
Switch agent backend from Anthropic to HF Inference Providers
Verified signatureUses smolagents InferenceClientModel with the logged-in user's own OAuth token (inference-api scope) instead of an Anthropic API key/Space secret.4b6e05e -
Add spaces package to requirements.txt
Verified signatured1c8c67 -
Satisfy ZeroGPU startup check without using GPU quota
Verified signatureAdd a no-op @spaces.GPU-decorated function so the ZeroGPU startup check passes; the agent itself never touches CUDA and this function is never invoked, so no GPU quota is spent.a245af6 -
Add smolagents + deps to requirements.txt
Verified signature7255773 -
Implement smolagents CodeAgent (Claude) for GAIA evaluation
Verified signatureReplaces the stub BasicAgent with a real agent using smolagents' CodeAgent, LiteLLMModel (Anthropic Claude), web search/wikipedia/webpage tools, and GAIA-style answer formatting. Also downloads per-question attached files and hands their local path to the agent.62e7e53
-
Update app.py
Verified signature81917a3 -
Update app.py
Verified signature0ee0419 -
Update app.py
Verified signaturee514fd7 -
Update app.py
Verified signatured59f015 -
Update requirements.txt
Verified signature1232637 -
Update app.py
Verified signature36ed51a -
Update app.py
Verified signature3db6293 -
Update app.py
Verified signaturec1fd3d2
-
Update app.py
Verified signatureb177367 -
Update app.py
Verified signature7d65c66 -
Update app.py
Verified signature9088b99 -
Update app.py
Verified signature3c4371f -
Update app.py
Verified signature9cde2c4 -
Update app.py
Verified signature3e5c3a6
-
Update app.py
Verified signature45d779e -
Update app.py
Verified signature1ca9f65 -
Update app.py
Verified signature4021bf3 -
Update app.py
Verified signaturef6f7832 -
Update app.py
Verified signatureb90251f -
Update app.py
Verified signature7e4a06b -
Update README.md
Verified signatured123508 -
Update app.py
Verified signature335f9f4 -
Update app.py
Verified signature31243f4 -
Update app.py
Verified signaturee80aab9 -
Update app.py
Verified signaturefa7d846