Add Tiny LLM weights, code and model card

README.md

@@ -0,0 +1,130 @@
+---
+license: mit
+language:
+  - en
+library_name: pytorch
+pipeline_tag: text-generation
+tags:
+  - tiny
+  - character-level
+  - gpt
+  - from-scratch
+  - educational
+---
+
+# Tiny LLM
+
+A very small GPT-style language model, trained from scratch in about 15 minutes
+on a laptop CPU. It writes one character at a time and only knows simple
+greetings: hello, how are you, what is your name, thanks, goodbye.
+
+It is a learning project, not a useful assistant. Anything that is not a
+greeting gets an unrelated greeting-style reply.
+
+**Try it in your browser:** [Tiny LLM demo Space](https://huggingface.co/spaces/Neural-Labs/NerualLabs)
+
+## Model details
+
+| | |
+| --- | --- |
+| Type | Decoder-only transformer (GPT-style), character level |
+| Parameters | 824,120 |
+| Layers | 4 |
+| Attention heads | 4 |
+| Embedding width | 128 |
+| Context length | 128 characters |
+| Vocabulary | 56 characters (the distinct characters in the training text) |
+| Language | English |
+| License | MIT |
+
+Each block uses pre-layer-norm, causal self-attention, and an MLP
+(128 → 512 → 128) with GELU. Token and position embeddings are both learned.
+
+## Files
+
+| File | What it is |
+| --- | --- |
+| `model.pt` | Trained weights: a dict with `state_dict`, `chars` (the vocabulary) and `config` |
+| `model.py` | The model code in PyTorch, about 60 lines |
+
+## How to use
+
+This is a custom PyTorch model, so it does not load with `transformers`.
+Download `model.py` and `model.pt`, then:
+
+```bash
+pip install torch huggingface_hub
+```
+
+```python
+import torch
+from huggingface_hub import hf_hub_download
+
+repo = "Neural-Labs/tiny-llm"
+hf_hub_download(repo, "model.py", local_dir=".")
+weights = hf_hub_download(repo, "model.pt")
+
+from model import TinyGPT
+
+ckpt = torch.load(weights, map_location="cpu", weights_only=True)
+chars = ckpt["chars"]
+stoi = {c: i for i, c in enumerate(chars)}
+model = TinyGPT(len(chars), **ckpt["config"])
+model.load_state_dict(ckpt["state_dict"])
+model.eval()
+
+
+def reply(message):
+    # Characters the model never saw are dropped.
+    clean = "".join(c for c in message.replace("\n", " ") if c in stoi).strip() or "hi"
+    prompt = f"User: {clean[:80]}\nBot:"
+    idx = torch.tensor([[stoi[c] for c in prompt]])
+    out = ""
+    for tok in model.generate(idx, max_new_tokens=100, temperature=0.5, top_k=5):
+        if chars[tok] == "\n":  # end of the bot's line
+            break
+        out += chars[tok]
+    return out.strip()
+
+
+print(reply("Hello!"))
+print(reply("How are you?"))
+```
+
+The prompt format must match the training data: `User: <message>\nBot:`.
+Generation stops at the first newline.
+
+## Training data
+
+12,000 generated exchanges in the form
+
+```
+User: <message>
+Bot: <reply>
+
+```
+
+built from a short hand-written list of greeting phrases and replies
+(hello, how are you, what is your name, thanks, goodbye, and similar).
+No external dataset was used.
+
+## Training
+
+| | |
+| --- | --- |
+| Steps | 4,000 |
+| Batch size | 32 |
+| Optimizer | AdamW |
+| Learning rate | 1e-3 with cosine decay |
+| Hardware | 4-core CPU |
+| Time | 14.5 minutes |
+| Final loss | train 0.165, validation 0.170 |
+
+## Limitations
+
+- It only knows greetings. Any other message gets a greeting-style reply that
+  has nothing to do with the question.
+- It works on characters, and drops any character that was not in its training
+  text (for example most non-English letters and emoji).
+- It has no memory of earlier messages: each reply sees only the latest message.
+- It is far too small to know facts, follow instructions, or reason.

model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a6b71ffe18697f2e8fd1afdbd8658ad4734a1d2c82d66f9353f8ecef78c71e87
+size 3311561

model.py

@@ -0,0 +1,66 @@
+"""A very small GPT-style language model (character level)."""
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class Block(nn.Module):
+    """One transformer block: self-attention, then a small feed-forward net."""
+
+    def __init__(self, n_embd, n_head, dropout):
+        super().__init__()
+        self.n_head = n_head
+        self.ln1 = nn.LayerNorm(n_embd)
+        self.qkv = nn.Linear(n_embd, 3 * n_embd)
+        self.proj = nn.Linear(n_embd, n_embd)
+        self.ln2 = nn.LayerNorm(n_embd)
+        self.mlp = nn.Sequential(
+            nn.Linear(n_embd, 4 * n_embd), nn.GELU(), nn.Linear(4 * n_embd, n_embd)
+        )
+        self.dropout = dropout
+
+    def forward(self, x):
+        B, T, C = x.shape
+        q, k, v = self.qkv(self.ln1(x)).split(C, dim=2)
+        q, k, v = (t.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) for t in (q, k, v))
+        # is_causal=True: each character may only look at the ones before it
+        a = F.scaled_dot_product_attention(
+            q, k, v, is_causal=True, dropout_p=self.dropout if self.training else 0.0
+        )
+        x = x + self.proj(a.transpose(1, 2).reshape(B, T, C))
+        return x + self.mlp(self.ln2(x))
+
+
+class TinyGPT(nn.Module):
+    def __init__(self, vocab_size, block_size=128, n_layer=4, n_head=4, n_embd=128, dropout=0.1):
+        super().__init__()
+        self.block_size = block_size
+        self.tok_emb = nn.Embedding(vocab_size, n_embd)
+        self.pos_emb = nn.Embedding(block_size, n_embd)
+        self.blocks = nn.ModuleList(Block(n_embd, n_head, dropout) for _ in range(n_layer))
+        self.ln_f = nn.LayerNorm(n_embd)
+        self.head = nn.Linear(n_embd, vocab_size)
+
+    def forward(self, idx, targets=None):
+        T = idx.shape[1]
+        x = self.tok_emb(idx) + self.pos_emb(torch.arange(T, device=idx.device))
+        for block in self.blocks:
+            x = block(x)
+        logits = self.head(self.ln_f(x))
+        loss = None
+        if targets is not None:
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
+        return logits, loss
+
+    @torch.no_grad()
+    def generate(self, idx, max_new_tokens, temperature=0.8, top_k=None):
+        """Yield one new token id at a time, feeding each back in as input."""
+        for _ in range(max_new_tokens):
+            logits, _ = self(idx[:, -self.block_size:])
+            logits = logits[:, -1, :] / max(temperature, 1e-5)
+            if top_k:
+                kth = torch.topk(logits, min(top_k, logits.size(-1))).values[:, [-1]]
+                logits[logits < kth] = -float("inf")
+            nxt = torch.multinomial(F.softmax(logits, dim=-1), 1)
+            idx = torch.cat([idx, nxt], dim=1)
+            yield nxt.item()