Upload README.md with huggingface_hub

README.md

@@ -8,46 +8,118 @@ tags: [hobbylm, mixture-of-experts, moe, sparse-moe]
 
 # HobbyLM-Computer-Use (500M MoE, GUI agent / tool use)
 
-Function-calling + accessibility-tree GUI-agent variant for computer-use tasks.
+HobbyLM-Computer-Use is the agentic variant: function calling plus a **text-only GUI agent** that reads a serialized accessibility tree (no pixels, no screenshots) and emits a grounded UI action. It can also decompose a multi-step goal and drive it to completion, deciding when it's `finish`ed.
 
-Part of the **HobbyLM** family — a from-scratch 500M sparse-MoE model trained on consumer-scale budgets.
+It's part of the **HobbyLM** family — a 500M sparse-MoE model (and its variants) built from scratch on a
+hobby budget: FineWeb, a handful of Modal H100 hours, a lot of ablations, and a from-scratch Rust engine
+([`hobby-rs`](https://github.com/harishsg993010/HobbyLM)) to run it on a laptop CPU.
+
+## Intended use
+
+Computer-use / GUI automation over a UI-Automation accessibility tree, and general tool / function calling. Serialize the screen as `SCREEN:\n[ControlType] "Name" (state) …`, give it the 12-action schema, and it returns a grounded action as JSON. Powers the Computer panel in the hobby-chat app.
 
 ## Architecture
 
-HobbyLM is a **sparse Mixture-of-Experts (MoE)** transformer (DeepSeek-V3 / Ling-style):
+Every HobbyLM variant shares one core: a **sparse Mixture-of-Experts (MoE)** decoder in the modern
+small-MoE style (DeepSeek-V3 / OLMoE lineage), where each design choice was picked by ablation rather
+than by guesswork.
 
 | Component | Value |
 |---|---|
-| Total parameters | ~500M (≈ a fraction active per token) |
-| Hidden size / layers | 768 / 16 (1 dense FFN layer, 15 MoE) |
+| Total parameters | ~500M (only a fraction is active per token) |
+| Hidden size / layers | 768 / 16 (first FFN dense, the rest MoE) |
 | Routed experts / active | 36 / top-6 (+ 1 always-on shared expert) |
-| Attention | GQA, 12 query / 3 KV heads, head-dim 128, per-head QK-norm |
-| Router | sigmoid gating, aux-loss-free balancing bias, no top-k renorm |
-| Positional | RoPE |
+| Attention | GQA, 12 query / 3 KV heads, decoupled head-dim 128, per-head QK-norm |
+| Router | sigmoid gating, DeepSeek-V3 aux-loss-free load balancing, no top-k renorm |
+| Positional | RoPE (θ up to 1e6 for the 8k-context checkpoints) |
 | Tokenizer | GPT-2 byte-level BPE (50,304 vocab, sentinel-padded) |
+| Optimizer | Muon on the 2-D + per-expert matrices, AdamW on everything else |
+
+The full ablation log (QK-norm is the single biggest lever; aux-loss-free beats classic aux-loss;
+≥32 experts and top-6 help; embedding-scaling hurt) lives in the project's architecture notes.
+
+## Benchmarks
+
+Held-out evaluation of the v4 checkpoint (accessibility-tree grounding + multi-step planning). `param-hallucination`
+is the rate of invented element names/arguments — strict tree-grounding in the data drives it to **0**.
 
+| Split | JSON-parse | Name-F1 | Value-acc | Exact-match | Param-halluc |
+|---|---|---|---|---|---|
+| Planning (multi-step goals) | 96.5% | 94.7% | — | 82.6% | 0.0% |
+| Grounding (real app trees) | ~96% | 95.5% | 91% | 78.4% | 0.0% |
+| Grounding (synthetic screens) | 100% | 90.7% | 88.6% | 72.5% | 0.0% |
 
-## Files
+For general (non-GUI) function calling, the HobbyLM tool-use lineage scores **~24% average on BFCL v3**
+(grammar-constrained) — strong relevance/abstention (relevance 77.8, beating the needle reference's 61.1),
+weaker on parallel multi-call, which is the 500M ceiling. Exact-match understates real quality: many "misses"
+are ambiguous numerics (e.g. *"give it a minute"* → `wait(60)` vs the reference `wait(7)`).
 
-- `model.safetensors` — the model weights (fp32).
-- `config.json` — architecture / hyperparameters.
-- GGUF builds (arch `hobbylm`) live in [`rootxhacker/HobbyLM-gguf`](https://huggingface.co/rootxhacker/HobbyLM-gguf).
+> **How these were measured.** All language-model scores are **0-shot** through our own port of
+> EleutherAI's `lm-evaluation-harness` (a custom `MoELMWrapper` that runs log-likelihood scoring over the
+> HobbyLM MoE + GPT-2 tokenizer). Reference models in the comparison table were run through the **identical
+> harness and task set**, so the numbers are apples-to-apples with ours — they are *not* copied from other
+> model cards. We validated the harness against published cards (e.g. TinyLlama 52.75 vs card 52.99). These
+> are small research models: read the numbers in context, not as leaderboard claims.
 
-## Loading (safetensors)
+## Usage
+
+### Python (PyTorch reference implementation)
+
+HobbyLM is a custom sparse-MoE architecture — there's no `transformers` `AutoModel` for it, so load it with
+the small reference implementation from the [GitHub repo](https://github.com/harishsg993010/HobbyLM):
 
 ```python
-import json, torch
+# HobbyLM is a CUSTOM sparse-MoE architecture, so load it with the reference implementation —
+# NOT transformers.AutoModelForCausalLM (there is no AutoModel mapping for this arch).
+# pip install torch safetensors tiktoken huggingface_hub
+# git clone https://github.com/harishsg993010/HobbyLM && cd HobbyLM
+
+import json, torch, tiktoken
+from huggingface_hub import hf_hub_download
 from safetensors.torch import load_file
-sd  = load_file("model.safetensors")
-cfg = json.load(open("config.json"))
-# rebuild the HobbyLM nn.Module from `cfg` and `load_state_dict(sd)`.
+from hobbylm.config import ModelConfig
+from hobbylm.model import MoETransformer
+from hobbylm.generate import generate
+
+repo = "rootxhacker/HobbyLM-Computer-Use"
+cfg = ModelConfig(**{k: v for k, v in json.load(open(hf_hub_download(repo, "config.json"))).items() if k != "preset"})
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+cfg.expert_backend = "grouped" if device.type == "cuda" else "bmm"
+
+model = MoETransformer(cfg).to(device).eval()
+model.load_state_dict(load_file(hf_hub_download(repo, "model.safetensors")))
+
+enc = tiktoken.get_encoding("gpt2")
+prompt = "USER: What is 7 plus 2?\nASSISTANT:"
+ids = torch.tensor([enc.encode_ordinary(prompt)], device=device)
+out = generate(model, ids, max_new_tokens=64, temperature=0.7, top_k=0, device=device,
+               repetition_penalty=1.3)               # temperature=0.0 for greedy
+print(enc.decode(out[0].tolist()))
 ```
 
-## Notes & limitations
+> For GUI / tool use, the real prompt format is `TOOLS: [<schema>]\nSCREEN:\n[ControlType] "Name" (state) …\nUSER: <instruction>\nASSISTANT:` and the model replies with a JSON action. The end-to-end agent loop lives in `agents/` in the repo.
+
+### GGUF + hobby-rs (CPU)
+
+GGUF builds (architecture `hobbylm`) live in [`rootxhacker/HobbyLM-gguf`](https://huggingface.co/rootxhacker/HobbyLM-gguf). They load
+directly in the from-scratch `hobby-rs` CPU engine — **stock llama.cpp won't load them** without registering
+the `hobbylm` architecture first.
+
+```bash
+hobby-rs --model HobbyLM-Computer-Use.gguf --prompt "..." --n 64
+```
+
+## Training
+
+Continue-SFT from the combined tool checkpoint on synthetic accessibility-tree data (Gemini-generated, strictly tree-validated) + real-app UI trees + planning trajectories, with a weighted loss. 13-action vocabulary (12 UI actions + `finish`).
+
+## Limitations
 
-- Research model at the ~500M scale: fluent but with the capability ceiling of a small model.
-- The GGUF uses a custom `hobbylm` architecture (see the GGUF repo) and needs `moe-rs` or a patched llama.cpp.
+- Per-step grounding is ~80% accurate; on **long** goals those errors compound (short tasks usually complete, long ones can drift) and there is no per-step recovery.
+- Trained on trees capped at ~45 elements (2k-context era); very large raw UI trees should be filtered.
+- Near-identical controls (e.g. digit buttons) occasionally mis-ground.
 
 ## License
 
-Apache-2.0.
+Apache-2.0. Weights aren't a substitute for judgement — this is a research / hobby model at the 500M scale,
+not a production system.