judotens commited on
Commit
aa9685e
·
verified ·
1 Parent(s): 16e0659

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +29 -26
README.md CHANGED
@@ -19,17 +19,17 @@ tags:
19
  pipeline_tag: text-generation
20
  ---
21
 
22
- # Flatbot-Mini-11M-Q4_0
23
 
24
  **Try it online:** https://chat.flatseek.io
25
 
26
- Flatbot-mini-11M-Q4_0 is the flagship compact conversational model for the **Flatseek** ecosystem.
27
 
28
- It is an **11 million parameter** decoder-only language model trained entirely from scratch using **FlatBuild**, exported to **GGUF**, quantized to **Q4_0**, and optimized for efficient CPU inference with **FlatRun**.
29
 
30
- The original FP32 checkpoint is approximately **135 MB**, while the **Q4_0 GGUF** version reduces the model size significantly, making it practical for lightweight local deployment without requiring a GPU.
31
 
32
- > **Experimental model:** Flatbot-mini-11M is intended for research, education, and experimentation. Although substantially more capable than Flatbot-micro-4M, it may still produce hallucinations, factual mistakes, or inconsistent reasoning.
33
 
34
  ---
35
 
@@ -44,7 +44,7 @@ The original FP32 checkpoint is approximately **135 MB**, while the **Q4_0 GGUF*
44
  | Attention | Grouped Query Attention (16 Query Heads / 4 KV Heads) |
45
  | Weight Tying | Yes |
46
  | Context Length | 512 |
47
- | Parameters | ~11M |
48
 
49
  ---
50
 
@@ -66,7 +66,7 @@ rope_theta = 10000
66
 
67
  # Training
68
 
69
- The model was trained entirely from random initialization using **FlatBuild**.
70
 
71
  ## Dataset
72
 
@@ -77,7 +77,7 @@ The model was trained entirely from random initialization using **FlatBuild**.
77
  | Validation Split | 5% |
78
  | Context Length | 512 tokens |
79
 
80
- The training dataset consists of approximately **10,000** conversational examples covering greetings, general knowledge, explanations, recommendations, question answering, follow-up conversations, and common assistant interactions.
81
 
82
  ---
83
 
@@ -98,25 +98,28 @@ The training dataset consists of approximately **10,000** conversational example
98
 
99
  # Quantization
100
 
101
- This release is provided as a **GGUF Q4_0** model optimized for lightweight inference.
102
 
103
  | Item | Value |
104
  |---|---:|
105
- | Original Format | FP32 |
106
  | Original Size | ~135 MB |
107
  | Quantization | GGUF Q4_0 |
108
- | Intended Runtime | FlatRun / llama.cpp compatible runtimes |
 
 
109
 
110
  ---
111
 
112
  # Features
113
 
114
  - Trained entirely from scratch
115
- - Custom tokenizer (1024 vocabulary)
 
116
  - Native chat template
117
  - RoPE positional embeddings
118
- - RMSNorm
119
- - SwiGLU
120
  - Grouped Query Attention (GQA)
121
  - Weight-tied embeddings
122
  - SafeTensors export
@@ -133,12 +136,12 @@ This release is provided as a **GGUF Q4_0** model optimized for lightweight infe
133
  ```bash
134
  pip install flatbuild
135
 
136
- flatbuild train configs/flatbot-mini-11M.yaml
137
 
138
  flatbuild export \
139
- outputs/flatbot-mini-11M-v2/*/checkpoint/final \
140
  --format gguf \
141
- --output flatbot-mini-11M
142
  ```
143
 
144
  ---
@@ -146,7 +149,7 @@ flatbuild export \
146
  ## LM Studio
147
 
148
  ```bash
149
- lms import flatbot-mini-11M-Q4_0.gguf
150
  ```
151
 
152
  ---
@@ -157,7 +160,7 @@ lms import flatbot-mini-11M-Q4_0.gguf
157
  pip install flatrun
158
 
159
  flatrun chat \
160
- --model flatbot-mini-11M-Q4_0.gguf \
161
  --temp 0.2
162
  ```
163
 
@@ -171,30 +174,30 @@ Loaded model in 0.03 s
171
  You: Who are you?
172
 
173
  Assistant:
174
- I'm Flatbot, a conversational AI assistant trained from scratch using the Flatseek ecosystem. I'm designed to answer questions, explain concepts, and assist with everyday tasks.
175
  ```
176
 
177
  ---
178
 
179
  # Purpose
180
 
181
- Flatbot-mini-11M demonstrates the complete **Flatseek AI pipeline**:
182
 
183
- 1. Build conversational datasets
184
  2. Train a tokenizer
185
- 3. Configure a transformer architecture
186
  4. Train entirely from scratch
187
  5. Export SafeTensors checkpoints
188
  6. Convert to GGUF
189
  7. Quantize to Q4_0
190
  8. Run efficient local inference with FlatRun
191
 
192
- The entire workflow is fully reproducible on consumer hardware and showcases how modern language models can be built without relying on pretrained foundation models.
193
 
194
  ---
195
 
196
  # Limitations
197
 
198
- Flatbot-mini-11M is still a relatively small language model compared with modern LLMs containing billions of parameters. While it offers noticeably stronger conversational ability than the earlier 4M demonstration model, it remains limited in factual knowledge, reasoning depth, multilingual performance, and long-context understanding.
199
 
200
- Its primary goal is to demonstrate an end-to-end open-source language model training, export, quantization, and inference pipeline built entirely within the Flatseek ecosystem.
 
19
  pipeline_tag: text-generation
20
  ---
21
 
22
+ # Flatbot-Mini-35M-Q4_0
23
 
24
  **Try it online:** https://chat.flatseek.io
25
 
26
+ Flatbot-Mini-35M-Q4_0 is the flagship compact conversational model of the **Flatseek** ecosystem.
27
 
28
+ It is a **34.9 million parameter** decoder-only Transformer trained entirely from scratch using **FlatBuild**, exported to **GGUF**, quantized to **Q4_0**, and optimized for fast CPU inference with **FlatRun**.
29
 
30
+ The original FP32 checkpoint is approximately **135 MB**, while the quantized **Q4_0 GGUF** model significantly reduces memory and storage requirements, making it well suited for efficient local deployment on consumer hardware.
31
 
32
+ > **Experimental model:** Flatbot-Mini-35M is designed for research, education, and experimentation. Although considerably more capable than Flatbot-Micro-4M, it may still generate hallucinations, factual inaccuracies, or inconsistent responses.
33
 
34
  ---
35
 
 
44
  | Attention | Grouped Query Attention (16 Query Heads / 4 KV Heads) |
45
  | Weight Tying | Yes |
46
  | Context Length | 512 |
47
+ | Parameters | 34.9M |
48
 
49
  ---
50
 
 
66
 
67
  # Training
68
 
69
+ Flatbot-Mini-35M was trained entirely from random initialization using **FlatBuild** without relying on pretrained weights.
70
 
71
  ## Dataset
72
 
 
77
  | Validation Split | 5% |
78
  | Context Length | 512 tokens |
79
 
80
+ The dataset contains approximately **10,000** multi-turn conversational examples covering greetings, question answering, explanations, recommendations, coding assistance, reasoning, and general-purpose assistant interactions.
81
 
82
  ---
83
 
 
98
 
99
  # Quantization
100
 
101
+ This release is distributed as a **GGUF Q4_0** model for efficient inference.
102
 
103
  | Item | Value |
104
  |---|---:|
105
+ | Original Format | FP32 SafeTensors |
106
  | Original Size | ~135 MB |
107
  | Quantization | GGUF Q4_0 |
108
+ | Parameters | 34.9M |
109
+ | Optimized For | CPU inference |
110
+ | Compatible Runtimes | FlatRun, llama.cpp, LM Studio, Ollama (GGUF) |
111
 
112
  ---
113
 
114
  # Features
115
 
116
  - Trained entirely from scratch
117
+ - 34.9M parameter Transformer
118
+ - Custom tokenizer (1,024 vocabulary)
119
  - Native chat template
120
  - RoPE positional embeddings
121
+ - RMSNorm normalization
122
+ - SwiGLU feed-forward network
123
  - Grouped Query Attention (GQA)
124
  - Weight-tied embeddings
125
  - SafeTensors export
 
136
  ```bash
137
  pip install flatbuild
138
 
139
+ flatbuild train configs/flatbot-mini-35M.yaml
140
 
141
  flatbuild export \
142
+ outputs/flatbot-mini-35M/*/checkpoint/final \
143
  --format gguf \
144
+ --output flatbot-mini-35M
145
  ```
146
 
147
  ---
 
149
  ## LM Studio
150
 
151
  ```bash
152
+ lms import flatbot-mini-35M-Q4_0.gguf
153
  ```
154
 
155
  ---
 
160
  pip install flatrun
161
 
162
  flatrun chat \
163
+ --model flatbot-mini-35M-Q4_0.gguf \
164
  --temp 0.2
165
  ```
166
 
 
174
  You: Who are you?
175
 
176
  Assistant:
177
+ I'm Flatbot, a conversational AI assistant trained from scratch using the Flatseek ecosystem. I'm here to answer questions, explain concepts, and help with everyday tasks.
178
  ```
179
 
180
  ---
181
 
182
  # Purpose
183
 
184
+ Flatbot-Mini-35M demonstrates the complete **Flatseek AI development pipeline**:
185
 
186
+ 1. Build a conversational dataset
187
  2. Train a tokenizer
188
+ 3. Configure a Transformer architecture
189
  4. Train entirely from scratch
190
  5. Export SafeTensors checkpoints
191
  6. Convert to GGUF
192
  7. Quantize to Q4_0
193
  8. Run efficient local inference with FlatRun
194
 
195
+ The entire workflow is fully reproducible on consumer hardware and showcases how modern language models can be trained, exported, quantized, and deployed without relying on proprietary foundation models.
196
 
197
  ---
198
 
199
  # Limitations
200
 
201
+ Flatbot-Mini-35M is still a compact language model compared with modern foundation models containing billions of parameters. While it offers substantially stronger conversational ability than the earlier Flatbot-Micro-4M demonstration model, it remains limited in factual knowledge, complex reasoning, multilingual capability, and long-context understanding.
202
 
203
+ Its primary purpose is to demonstrate an end-to-end open-source training, quantization, and inference pipeline built entirely within the Flatseek ecosystem.