Upload folder using huggingface_hub

README.md

@@ -0,0 +1,45 @@
+---
+library_name: transformers
+tags:
+  - minigpt
+  - custom_code
+  - text-generation
+license: mit
+---
+
+# tinystories-50m-instruct
+
+A ~50M-parameter GPT-style transformer (6 layers, 12 heads, 768 embd, 1024
+context) trained from scratch on TinyStories and instruction-fine-tuned to
+chat in a `User: ... / Assistant: ...` format.
+
+## Quick start
+
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+tokenizer = AutoTokenizer.from_pretrained("hayder86al/tinystories-50m-instruct")
+model = AutoModelForCausalLM.from_pretrained("hayder86al/tinystories-50m-instruct", trust_remote_code=True)
+
+def chat(prompt):
+    text = f"User: {prompt}\nAssistant:"
+    inputs = tokenizer(text, return_tensors="pt")
+    outputs = model.generate(
+        **inputs, max_new_tokens=120, do_sample=True, temperature=0.5,
+        top_k=40, use_cache=False,
+        pad_token_id=tokenizer.pad_token_id, eos_token_id=tokenizer.eos_token_id,
+    )
+    text_out = tokenizer.decode(outputs[0], skip_special_tokens=True)
+    return text_out.split("Assistant:")[-1].strip()
+
+print(chat("What is 2+2?"))
+```
+
+`trust_remote_code=True` is required because this repo defines a custom
+architecture (`modeling_minigpt.py` / `configuration_minigpt.py`) — review
+that code before running it.
+
+## Limitations
+
+Single-turn only (no conversation memory), 1024-token context, small model —
+expect simple, sometimes ungrammatical answers, not factual accuracy.

config.json

@@ -0,0 +1,20 @@
+{
+  "architectures": [
+    "MiniGPTForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_minigpt.MiniGPTConfig",
+    "AutoModelForCausalLM": "modeling_minigpt.MiniGPTForCausalLM"
+  },
+  "context_len": 1024,
+  "dropout": 0.1,
+  "dtype": "float32",
+  "eos_token_id": 1,
+  "model_type": "minigpt",
+  "n_embd": 768,
+  "n_head": 12,
+  "n_layer": 6,
+  "pad_token_id": 0,
+  "transformers_version": "5.16.1",
+  "vocab_size": 8000
+}

configuration_minigpt.py

@@ -0,0 +1,27 @@
+"""HF config wrapper for MiniGPT (train_50M.py's CFG, made introspectable)."""
+
+from transformers import PretrainedConfig
+
+
+class MiniGPTConfig(PretrainedConfig):
+    model_type = "minigpt"
+
+    def __init__(
+        self,
+        vocab_size: int = 8_000,
+        n_embd: int = 768,
+        n_head: int = 12,
+        n_layer: int = 6,
+        context_len: int = 1024,
+        dropout: float = 0.1,
+        pad_token_id: int = 0,
+        eos_token_id: int = 1,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.n_embd = n_embd
+        self.n_head = n_head
+        self.n_layer = n_layer
+        self.context_len = context_len
+        self.dropout = dropout
+        super().__init__(pad_token_id=pad_token_id, eos_token_id=eos_token_id, **kwargs)

generation_config.json

@@ -0,0 +1,8 @@
+{
+  "_from_model_config": true,
+  "eos_token_id": 1,
+  "output_attentions": false,
+  "output_hidden_states": false,
+  "pad_token_id": 0,
+  "transformers_version": "5.16.1"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c364d1109362057112581a4d8832f193981d794cb00158494b7fc5bc03b264ae
+size 222345008

modeling_minigpt.py

@@ -0,0 +1,120 @@
+"""HF-compatible MiniGPT — same architecture as train_50M.MiniGPT, wrapped as a
+PreTrainedModel so it works with AutoModelForCausalLM / .generate() /
+push_to_hub() (requires trust_remote_code=True, since this is a custom arch).
+
+No KV cache, learned absolute positional embeddings: every forward re-crops the
+sequence to the last `context_len` tokens and re-embeds positions from 0 —
+that mirrors train_50M/chat.py exactly and is a correctness requirement, not
+an optimization.
+"""
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import GenerationMixin, PreTrainedModel
+from transformers.modeling_outputs import CausalLMOutput
+
+from .configuration_minigpt import MiniGPTConfig
+
+
+class CausalSelfAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        assert config.n_embd % config.n_head == 0
+        self.n_head = config.n_head
+        self.head_dim = config.n_embd // config.n_head
+        self.qkv = nn.Linear(config.n_embd, 3 * config.n_embd, bias=False)
+        self.proj = nn.Linear(config.n_embd, config.n_embd, bias=False)
+        self.dropout = config.dropout
+        self.resid_drop = nn.Dropout(config.dropout)
+
+    def forward(self, x):
+        B, T, C = x.shape
+        q, k, v = self.qkv(x).split(C, dim=2)
+
+        def split_heads(t):
+            return t.view(B, T, self.n_head, self.head_dim).transpose(1, 2)
+
+        q, k, v = split_heads(q), split_heads(k), split_heads(v)
+        out = F.scaled_dot_product_attention(
+            q, k, v, is_causal=True,
+            dropout_p=self.dropout if self.training else 0.0,
+        )
+        out = out.transpose(1, 2).contiguous().view(B, T, C)
+        return self.resid_drop(self.proj(out))
+
+
+class FeedForward(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.fc = nn.Linear(config.n_embd, 4 * config.n_embd)
+        self.act = nn.GELU()
+        self.proj = nn.Linear(4 * config.n_embd, config.n_embd)
+        self.drop = nn.Dropout(config.dropout)
+
+    def forward(self, x):
+        return self.drop(self.proj(self.act(self.fc(x))))
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.ln1 = nn.LayerNorm(config.n_embd)
+        self.attn = CausalSelfAttention(config)
+        self.ln2 = nn.LayerNorm(config.n_embd)
+        self.ff = FeedForward(config)
+
+    def forward(self, x):
+        x = x + self.attn(self.ln1(x))
+        x = x + self.ff(self.ln2(x))
+        return x
+
+
+class MiniGPTForCausalLM(PreTrainedModel, GenerationMixin):
+    config_class = MiniGPTConfig
+    base_model_prefix = "minigpt"
+    main_input_name = "input_ids"
+
+    def __init__(self, config: MiniGPTConfig):
+        super().__init__(config)
+        self.token_emb = nn.Embedding(config.vocab_size, config.n_embd)
+        self.pos_emb = nn.Embedding(config.context_len, config.n_embd)
+        self.drop = nn.Dropout(config.dropout)
+        self.blocks = nn.Sequential(*[TransformerBlock(config) for _ in range(config.n_layer)])
+        self.ln_f = nn.LayerNorm(config.n_embd)
+        self.head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+        self.head.weight = self.token_emb.weight   # weight tying, as in train_50M.MiniGPT
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.token_emb
+
+    def set_input_embeddings(self, value):
+        self.token_emb = value
+
+    def get_output_embeddings(self):
+        return self.head
+
+    def set_output_embeddings(self, value):
+        self.head = value
+
+    def forward(self, input_ids=None, attention_mask=None, labels=None, **kwargs):
+        ctx = self.config.context_len
+        idx = input_ids[:, -ctx:]
+        pos = torch.arange(idx.size(1), device=idx.device)
+        x = self.drop(self.token_emb(idx) + self.pos_emb(pos))
+        x = self.blocks(x)
+        x = self.ln_f(x)
+        logits = self.head(x)
+
+        loss = None
+        if labels is not None:
+            labels = labels[:, -ctx:]
+            loss = F.cross_entropy(
+                logits.view(-1, logits.size(-1)), labels.view(-1), ignore_index=-100)
+        return CausalLMOutput(loss=loss, logits=logits)
+
+    # No KV cache exists for this architecture — every generation step resends
+    # the (cropped) full sequence, so there is nothing to thread through here.
+    def prepare_inputs_for_generation(self, input_ids, **kwargs):
+        return {"input_ids": input_ids}

tokenizer_config.json

@@ -0,0 +1,8 @@
+{
+  "backend": "tokenizers",
+  "eos_token": "<eos>",
+  "model_max_length": 1024,
+  "pad_token": "<pad>",
+  "tokenizer_class": "TokenizersBackend",
+  "unk_token": "<unk>"
+}