Brand v2: banner, eyebrow, About footer, 'Where it falls short'; Udgam Labs name, lowercase ids

README.md

@@ -4,8 +4,9 @@ language:
 - en
 library_name: mlx
 pipeline_tag: text-generation
-base_model: UdgamLabs/Udgam-384M-Chat
+base_model: udgamlabs/Udgam-384M-Chat
 tags:
+- udgamlabs
 - gemma4
 - tool-calling
 - function-calling
@@ -17,15 +18,19 @@ tags:
 base_model_relation: quantized
 ---
 
+![Udgam-384M-Chat-MLX · Udgam Labs](./banner.png)
+
+**UDGAM LABS** · Tool-calling model, MLX · 384M parameters
+
 # Udgam-384M-Chat-MLX
 
-**This repo holds the MLX fp16 version of [Udgam-384M-Chat](https://huggingface.co/UdgamLabs/Udgam-384M-Chat)** for Apple silicon:
-`pip install mlx-lm`, then `mlx_lm.generate --model UdgamLabs/Udgam-384M-Chat-MLX --prompt "..."` (or `mlx_lm.server`
+**This repo holds the MLX fp16 version of [Udgam-384M-Chat](https://huggingface.co/udgamlabs/Udgam-384M-Chat)** for Apple silicon:
+`pip install mlx-lm`, then `mlx_lm.generate --model udgamlabs/Udgam-384M-Chat-MLX --prompt "..."` (or `mlx_lm.server`
 for an OpenAI-compatible API). Converted with mlx-lm 0.32. The rest of this page is the main model card.
 
 **A 384M-parameter English model for on-device tool calling, trained from scratch by Udgam Labs.** It is a
 decoder-only transformer compatible with the Gemma 4 architecture (`Gemma4ForCausalLM`), so it runs unmodified in
-Ollama, LM Studio, llama.cpp, MLX and 🤗 transformers. Context: 4,096 tokens. The 4-bit GGUF is 238 MB. Built for four
+Ollama, LM Studio, llama.cpp, MLX and transformers. Context: 4,096 tokens. The 4-bit GGUF is 238 MB. Built for four
 jobs inside an app:
 
 1. **App actions**: turn a request into a function call. "wake me at 6:45, label it gym" →
@@ -39,7 +44,7 @@ jobs inside an app:
 **What it is not:** a knowledgeable general assistant (at 384M parameters it makes factual mistakes), or a
 safety-hardened open chatbot. Use it where *your app* decides which tools exist and confirms risky actions.
 
-**Model family:** [Udgam-384M-Base](https://huggingface.co/UdgamLabs/Udgam-384M-Base) (2,048 context) · [Udgam-384M-Base-4K](https://huggingface.co/UdgamLabs/Udgam-384M-Base-4K) · [Udgam-384M-Chat](https://huggingface.co/UdgamLabs/Udgam-384M-Chat) · [GGUF](https://huggingface.co/UdgamLabs/Udgam-384M-Chat-GGUF) · [MLX](https://huggingface.co/UdgamLabs/Udgam-384M-Chat-MLX) · [MLX 4-bit](https://huggingface.co/UdgamLabs/Udgam-384M-Chat-MLX-4bit) · **[Technical report](https://udgamlabs.com/reports/udgam-384m)**
+**Model family:** [Udgam-384M-Base](https://huggingface.co/udgamlabs/Udgam-384M-Base) (2,048 context) · [Udgam-384M-Base-4K](https://huggingface.co/udgamlabs/Udgam-384M-Base-4K) · [Udgam-384M-Chat](https://huggingface.co/udgamlabs/Udgam-384M-Chat) · [GGUF](https://huggingface.co/udgamlabs/Udgam-384M-Chat-GGUF) · [MLX](https://huggingface.co/udgamlabs/Udgam-384M-Chat-MLX) · [MLX 4-bit](https://huggingface.co/udgamlabs/Udgam-384M-Chat-MLX-4bit) · **[Technical report](https://udgamlabs.com/reports/udgam-384m)**
 
 | | |
 |---|---|
@@ -55,7 +60,7 @@ safety-hardened open chatbot. Use it where *your app* decides which tools exist
 
 ### Ollama (recommended)
 
-Download a GGUF from the [GGUF repo](https://huggingface.co/UdgamLabs/Udgam-384M-Chat-GGUF) and create a model with its `Modelfile` (it uses
+Download a GGUF from the [GGUF repo](https://huggingface.co/udgamlabs/Udgam-384M-Chat-GGUF) and create a model with its `Modelfile` (it uses
 Ollama's native Gemma 4 renderer and parser, so tool calls and the optional thinking mode work as trained):
 
 ```
@@ -91,14 +96,14 @@ For this model thinking is optional and usually not better (see the tables).
 
 ### LM Studio
 
-1. Search for `UdgamLabs/Udgam-384M-Chat-GGUF` (or drop the `.gguf` into LM Studio's models folder) and pick `Q8_0`
+1. Search for `udgamlabs/Udgam-384M-Chat-GGUF` (or drop the `.gguf` into LM Studio's models folder) and pick `Q8_0`
    (default) or `Q4_K_M` (phones / low memory).
 2. The GGUF carries the Gemma 4 chat template, so the prompt format is set automatically. Set **Context Length 4096**,
    temperature 0.1, top-p 0.9, top-k 40, repeat penalty 1.1.
 3. For tool use, call LM Studio's OpenAI-compatible server (`/v1/chat/completions` with `tools`). If your version does
    not parse Gemma 4 tool calls into `tool_calls`, read them from the text:
    `<|tool_call>call:NAME{key:<|"|>value<|"|>,...}<tool_call|>`.
-4. **Using this MLX version in LM Studio:** before loading, open the model's load settings (⚙️, or turn on
+4. **Using this MLX version in LM Studio:** before loading, open the model's load settings (the gear icon, or turn on
    "Manually choose model load parameters") and set **Max Concurrent Predictions to 1**. With the default (parallel)
    setting, LM Studio's MLX engine (tested with mlx-llm 1.13.1) fails on text-only Gemma 4 models with
    `KeyError: 'text_config'`. The GGUF version needs no change.
@@ -119,7 +124,7 @@ build d7a695e).
 
 ```bash
 pip install mlx-lm
-mlx_lm.generate --model UdgamLabs/Udgam-384M-Chat-MLX-4bit --prompt "Give me three tips for sleeping better." --max-tokens 200
+mlx_lm.generate --model udgamlabs/Udgam-384M-Chat-MLX-4bit --prompt "Give me three tips for sleeping better." --max-tokens 200
 ```
 
 In **LM Studio**, load this model with **Max Concurrent Predictions = 1** (see the LM Studio section above).
@@ -129,8 +134,8 @@ In **LM Studio**, load this model with **Max Concurrent Predictions = 1** (see t
 ```python
 # tested with transformers 5.18
 from transformers import AutoTokenizer, AutoModelForCausalLM
-tok = AutoTokenizer.from_pretrained("UdgamLabs/Udgam-384M-Chat")
-model = AutoModelForCausalLM.from_pretrained("UdgamLabs/Udgam-384M-Chat")
+tok = AutoTokenizer.from_pretrained("udgamlabs/Udgam-384M-Chat")
+model = AutoModelForCausalLM.from_pretrained("udgamlabs/Udgam-384M-Chat")
 tools = [{"type": "function", "function": {"name": "set_alarm", "description": "Set an alarm at a local time.",
           "parameters": {"type": "object", "properties": {"time": {"type": "string", "description": "HH:MM 24h"},
                          "label": {"type": "string"}}, "required": ["time"]}}}]
@@ -238,11 +243,6 @@ RAM = macOS peak memory footprint of the CPU run. Phones are slower; the relativ
 document → JSON against an unseen schema (schema-valid 56–60% vs 24–38%) and shell behaviour (86% vs 59–67%). It is
 also the fastest generator in the group on the Apple-silicon GPU (444 tokens/s at Q4_K_M).
 
-**Where it is behind:** exact tool calls and instruction following. Both Granite 4.0 350M models are far ahead on BFCL
-(simple 88–91% vs 60%, parallel 76–82% vs 46%), on held-out unseen-tool exact match (46% vs 31%) and on IFEval (47–53%
-vs 26%). Qwen3-0.6B (about 1.6 times the parameters) is also ahead on BFCL and IFEval. Shell commands are rarely *correct*
-when executed for any model here (0–8%).
-
 **Fairness notes:**
 - The held-out task sets use this model's prompt conventions (date/device system headers, normalised argument values
   such as ISO dates); other models lose some points to convention rather than ability. BFCL, IFEval and XSTest are
@@ -282,6 +282,15 @@ when executed for any model here (0–8%).
 Perplexity on held-out chat and tool-use text: f16 3.193, Q8_0 +0.07%, Q6_K +0.5%, Q4_K_M (imatrix) +4.2%, Q4_0
 (imatrix) +8.3%.
 
+## Where it falls short
+
+It is behind on exact tool calls and instruction following. Both Granite 4.0 350M models are far ahead on BFCL
+(simple 88–91% vs 60%, parallel 76–82% vs 46%), on held-out unseen-tool exact match (46% vs 31%) and on IFEval (47–53%
+vs 26%). Qwen3-0.6B (about 1.6 times the parameters) is also ahead on BFCL and IFEval. Shell commands are rarely *correct*
+when executed for any model here (0–8%).
+
+On the neutral BFCL irrelevance test, Qwen3-0.6B and SmolLM2 also beat it (see the fairness notes above). For safety results and the tasks it is not meant for, see *Safety* and *Intended use and limitations*.
+
 ## Safety
 
 XSTest (200 clearly unsafe + 250 safe-but-alarming prompts), greedy, thinking off: **the model declines 54% of clearly harmful prompts and over-refuses 15% of safe ones**
@@ -315,7 +324,7 @@ email → JSON with a given schema; suggesting shell commands for a person to re
 ## Training
 
 Pretrained from scratch on roughly 8B tokens of licensed public and synthetic text (base:
-[Udgam-384M-Base-4K](https://huggingface.co/UdgamLabs/Udgam-384M-Base-4K)), then post-trained with supervised fine-tuning for chat and tool use.
+[Udgam-384M-Base-4K](https://huggingface.co/udgamlabs/Udgam-384M-Base-4K)), then post-trained with supervised fine-tuning for chat and tool use.
 See the [technical report](https://udgamlabs.com/reports/udgam-384m) for an overview.
 
 ## Licence
@@ -329,7 +338,7 @@ Apache-2.0 (weights and tokenizer); see `LICENSE`. The chat template is the offi
   title        = {Udgam-384M: a small on-device tool-calling model trained from scratch},
   author       = {{Udgam Labs}},
   year         = {2026},
-  howpublished = {\url{https://huggingface.co/UdgamLabs/Udgam-384M-Chat}},
+  howpublished = {\url{https://huggingface.co/udgamlabs/Udgam-384M-Chat}},
   note         = {Technical report: \url{https://udgamlabs.com/reports/udgam-384m}}
 }
 ```
@@ -337,3 +346,7 @@ Apache-2.0 (weights and tokenizer); see `LICENSE`. The chat template is the offi
 ## Contact
 
 Udgam Labs. Please use the Community (discussions) tab of the model page for questions, issues and feedback.
+
+---
+
+**About Udgam Labs.** Udgam Labs is an independent AI research lab. We work from first principles: before we build, we understand why things work, whether we are training a model, fine-tuning one or applying AI to a new problem. Udgam (उद्गम) is Sanskrit for origin. [udgamlabs.com](https://udgamlabs.com)