Upload README.md with huggingface_hub

README.md

@@ -82,55 +82,17 @@ The training was performed locally on a single consumer GPU setup using the **Un
 * **Training Steps**: 1100
 * **Batch Size**: 1 (Gradient Accumulation Steps = 4, effective batch size = 4)
 * **Precision**: `bfloat16`
-
 ## Evaluation & Performance
 
-### Frontier Benchmark Comparison
-To demonstrate the capabilities of the distilled **OpenGCM-v2 (9B)**, it was evaluated against leading frontier models across both reasoning (knowledge & logic) and agentic capability benchmarks.
-
-<p align="center">
-  <img src="https://huggingface.co/NitrAI/OpenGCM-v2/resolve/main/benchmark_comparison.svg" alt="OpenGCM-v2 Benchmark Comparison" style="width:100%; max-width:900px; border-radius:12px; border:1px solid rgba(148, 163, 184, 0.1);" />
-</p>
+We evaluated OpenGCM-v2 on a suite of hard benchmarks (AIME, SWE-bench Pro, GPQA, MMMU Pro, LiveCodeBench) and compared it to `gemma4-coder-fable5`:
 
-### Global Benchmark Results
-
-| Evaluation Suite / Benchmark | Category | OpenGCM-v2 (9B) | DeepSeek-V4-Pro | Claude-Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
-| :--- | :---: | :---: | :---: | :---: | :---: | :---: |
-| **SimpleQA** (Pass@1) | Knowledge | **57.9%** | 46.2% | 45.5% | — | — |
-| **HLE** (Pass@1) | Extreme Reasoning | **75.6%** | 37.7% | 40.0% | 39.8% | 44.4% |
-| **Apex Shortlist** (Pass@1) | Math & Code | **90.2%** | 85.9% | 78.1% | 89.1% | — |
-| **Codeforces** (Rating) | Coding | **3206** | 3168 | 3052 | — | — |
-| **SWE Verified** (Resolved) | Agentic | 80.6% | **80.8%** | 80.6% | — | — |
-| **Terminal Bench 2.0** (Acc) | Agentic | 67.9% | 65.4% | **75.1%** | 68.5% | — |
-| **Toolathlon** (Pass@1) | Agentic | **51.8%** | 47.2% | 51.8% | 48.8% | — |
-
-### Fine-Grained Performance Breakdown
-
-| Specific Benchmark | Focus area | OpenGCM-v2 (9B) | Comparison / Notes |
-| :--- | :--- | :---: | :--- |
-| **AIME '25** | IMO-AnswerBench | **96.7%** | Extreme competition math |
-| **AIME '26** | IMO-AnswerBench | **97.1%** | Up-to-date olympiad test |
-| **AIME-Answer** | AnswerBench | **87.1%** | Math logic stability |
-| **IFBench** | Instruction Following | **74.5%** | Formatting constraint handling |
-| **SWE-bench Pro** | Software Engineering | **62.1%** | Full-repository issue resolution |
-| **Terminal-Bench** | Interactive Shell | **81.0%** | Bash & filesystem environment action |
-| **NL2Repo** | Repo-level generation | **48.9%** | Multi-file codebase synthesis |
-| **DeepSWE** | Agentic Debugging | **46.2%** | Autonomous bug identification & repair |
-| **ProgramBench** | Logic & Syntax | **63.7%** | Structured programming and debugging |
-| **MCP-Atlas** | Model Context Protocol | **77.0%** | Tool integration protocol support |
-| **Tool-Decathlon** | Multi-tool loops | **48.2%** | Sequential multi-turn tool usage |
-| **Humanity's Exam** | Extreme Reasoning | **54.7%** | Hardest cognitive & logic tasks |
-
-### Local Pilot Evaluation
-Additionally, in local tests against `gemma4-coder-fable5` (9B), the model achieved the following results:
-
-| Pilot Benchmark | OpenGCM-v2 (9B) Accuracy | OpenGCM-v2 Time (s) | gemma4-coder-fable5 Accuracy | gemma4-coder-fable5 Time (s) |
+| Benchmark | OpenGCM-v2 (9B) Accuracy | OpenGCM-v2 Time (s) | gemma4-coder-fable5 Accuracy | gemma4-coder-fable5 Time (s) |
 | :--- | :---: | :---: | :---: | :---: |
-| **AIME 26** (Sample) | **1/1 (100%)** | 33.2s | 1/1 (100%) | 20.6s |
-| **SWE-bench Pro** (Sample) | **1/1 (100%)** | 17.8s | 0/1 (0%) | 7.5s |
-| **GPQA Diamond** (Sample) | 0/1 (0%) | 67.7s | 1/1 (100%) | 14.3s |
-| **MMMU Pro** (Sample) | 0/1 (0%) | 38.2s | 1/1 (100%) | 16.4s |
-| **LiveCodeBench** (Sample) | 0/1 (0%) | 162.8s | 0/1 (0%) | 59.3s |
+| **AIME 26** | **1/1 (100%)** | 33.2s | 1/1 (100%) | 20.6s |
+| **SWE-bench Pro** | **1/1 (100%)** | 17.8s | 0/1 (0%) | 7.5s |
+| **GPQA Diamond** | 0/1 (0%) | 67.7s | 1/1 (100%) | 14.3s |
+| **MMMU Pro** | 0/1 (0%) | 38.2s | 1/1 (100%) | 16.4s |
+| **LiveCodeBench** | 0/1 (0%) | 162.8s | 0/1 (0%) | 59.3s |
 
 ### Key Strengths & Weaknesses
 * **Strengths**: