Update instruction checkpoint with verified BF16 weights

Updated SFT hyperparameters, unchanged training data. BF16 conversion and vLLM loading/generation validated for this checkpoint.

README.md

@@ -32,9 +32,9 @@ For the base (pretrained) model, see
 
 ## Chat template
 
-The Llama-2-style chat template ships in `tokenizer_config.json` — apply it
-with the tokenizer. The BOS token must come from the tokenizer, **not** as a
-literal `"<s>"` string in your prompt text.
+The Llama-2-style chat template ships in `chat_template.jinja`. Apply it
+with the tokenizer. The template includes BOS/EOS tokens, so use
+`add_special_tokens=False` when tokenizing the rendered prompt.
 
 ```python
 import torch
@@ -48,7 +48,7 @@ prompt = tokenizer.apply_chat_template(
     [{"role": "user", "content": "What is the capital of France?"}],
     tokenize=False,
 )
-inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
+inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).to(model.device)
 output = model.generate(**inputs, max_new_tokens=128, do_sample=True,
                         temperature=0.7, top_p=0.95, use_cache=True,
                         eos_token_id=tokenizer.eos_token_id,
@@ -60,3 +60,11 @@ print(tokenizer.decode(output[0, inputs["input_ids"].shape[-1]:], skip_special_t
 
 - Knowledge limited to the 2014 data vintage.
 - No RLHF or safety tuning; outputs can be confidently wrong.
+
+## Checkpoint update (2026-10-14)
+
+This checkpoint uses the updated SFT hyperparameters with the same training data.
+The FP32 checkpoint was converted to BF16; saved tensors were verified against
+the expected BF16 casts. Loading and a short generation test passed with
+vLLM 0.27.1. This is a functional smoke test, not a quality benchmark.
+See `training_recipe.json` for the training configuration.

chat_template.jinja

@@ -0,0 +1,5 @@
+{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% else %}{% set loop_messages = messages %}{% endif %}{% for message in loop_messages %}{% if loop.index0 == 0 and system_message is defined %}{% set content = '<<SYS>>
+' + system_message + '
+<</SYS>>
+
+' + message['content'] %}{% else %}{% set content = message['content'] %}{% endif %}{% if message['role'] == 'user' %}{{ '<s>' + '[INST] ' + content + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ content + '</s>' }}{% endif %}{% endfor %}

config.json

@@ -18,14 +18,12 @@
   "num_attention_heads": 16,
   "num_hidden_layers": 24,
   "num_key_value_heads": 16,
-  "pad_token_id": 2,
   "pretraining_tp": 1,
   "rms_norm_eps": 1e-05,
   "rope_scaling": null,
   "rope_theta": 10000.0,
   "tie_word_embeddings": false,
-  "torch_dtype": "bfloat16",
-  "transformers_version": "4.51.0",
+  "transformers_version": "4.57.6",
   "use_cache": false,
   "vocab_size": 32000
 }

generation_config.json

@@ -1,9 +1,6 @@
 {
   "_from_model_config": true,
   "bos_token_id": 1,
-  "eos_token_id": [
-    2
-  ],
-  "pad_token_id": 2,
-  "transformers_version": "4.51.0"
+  "eos_token_id": 2,
+  "transformers_version": "4.57.6"
 }

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:fe4a1c0b581fb2a620497a2c59a64259b347d05fd91c6b99d106a0342c32ee8a
+oid sha256:37f407cc7b6e913ab94807ab75b18b29343893327e236b4dbcb4842675c6984a
 size 2690871976

tokenizer_config.json

@@ -29,7 +29,6 @@
     }
   },
   "bos_token": "<s>",
-  "chat_template": "{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% else %}{% set loop_messages = messages %}{% endif %}{% for message in loop_messages %}{% if loop.index0 == 0 and system_message is defined %}{% set content = '<<SYS>>\n' + system_message + '\n<</SYS>>\n\n' + message['content'] %}{% else %}{% set content = message['content'] %}{% endif %}{% if message['role'] == 'user' %}{{ '<s>' + '[INST] ' + content + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ content + '</s>' }}{% endif %}{% endfor %}",
   "clean_up_tokenization_spaces": false,
   "eos_token": "</s>",
   "extra_special_tokens": {},

training_recipe.json

@@ -0,0 +1,15 @@
+{
+  "base_model": "/data/users/zhenyugao/sjd/last_ckpt/ckpt_model_single_2014_iter_0025000/hf_model",
+  "tokenizer": "/data/users/zhenyugao/sjd/LLaMA-Factory/saves/dated_mining_2013_final_v4_financialv2",
+  "data": "/data/users/zhenyugao/sjd/LLaMA-Factory/data/sft_mixed_pt10tok_2014_with_nyt2k.json",
+  "loss": "model_native",
+  "world_size": 1,
+  "per_device_batch": 1,
+  "gradient_accumulation": 32,
+  "global_batch": 32,
+  "learning_rate": 1e-05,
+  "epochs": 3,
+  "total_steps": 5460,
+  "max_length": 2048,
+  "seed": 42
+}