Upload README.md with huggingface_hub

README.md

@@ -56,18 +56,17 @@ uv run yc-bench run --model <your-model> --seed 3 --config medium --no-live
 
 Any [LiteLLM-compatible](https://docs.litellm.ai/docs/providers) model string works (e.g. `openai/gpt-5.2`, `anthropic/claude-sonnet-4-6`, `openrouter/google/gemini-3-flash-preview`).
 
-Results are saved to `results/` as JSON. Each result contains `terminal_reason` (horizon_end or bankruptcy) and `final_funds_cents`.
+Results are saved to `results/` as JSON.
 
 ## Scoring
 
-```
-composite = 0.5 * survival + 0.5 * normalised_funds
-```
+Report the **average final funds (in USD) across seeds 1, 2, 3**. Agents that go bankrupt report $0.
 
-- **survival**: 1 if the agent reaches the 1-year horizon without bankruptcy, 0 otherwise
-- **normalised_funds**: `min(log(1 + final_funds / initial_funds) / log(101), 1.0)` — 0 if bankrupt, ~0.15 at breakeven, 1.0 at 100x return
+Read `final_funds_cents` from the result JSON (or query `SELECT funds_cents FROM companies` in the SQLite DB). Negative values (bankruptcy) count as $0.
 
-Report the **average composite score across seeds 1, 2, 3**.
+```
+score = average(max(0, final_funds_cents / 100) for each seed)
+```
 
 ## Submitting Results
 
@@ -77,12 +76,12 @@ Add a `.eval_results/yc-bench.yaml` file to your model's HF repo:
 - dataset:
     id: collinear-ai/yc-bench
     task_id: medium
-  value: 0.65
+  value: 12500
   date: "2025-03-20"
   source:
     url: https://huggingface.co/<your-model>
     name: "Self-reported"
-  notes: "seeds 1,2,3 averaged"
+  notes: "avg final funds (USD) across seeds 1,2,3"
 ```
 
 ## License