You can deploy this model and start earning money today!

devsatish/
fixed_model

Name: fixed_model
Rating: 2.5 (6 reviews)

meta-llama/Llama-3.1-8B

finetune

endpoints_compatible

Views

No views yet

Fixed LLaMA 3.1 8B Configuration

This repository contains corrected configuration files for Meta-Llama-3.1-8B, fixing issues from the original yunmorning/broken-model.

Problem Identified

The original model's config.json contained Qwen3 model settings while the README claimed it was based on Meta-Llama-3.1-8B. This mismatch causes inference failures because:

The architecture class (Qwen3ForCausalLM) doesn't match the actual model weights
Token IDs (BOS, EOS) are wrong for LLaMA tokenizer
Vocabulary size mismatches the LLaMA tokenizer
Layer count and intermediate size are incorrect for LLaMA 3.1 8B

Changes Made

config.json

Field	Original (Broken)	Fixed	Reason
`architectures`	`["Qwen3ForCausalLM"]`	`["LlamaForCausalLM"]`	Must match LLaMA model class
`model_type`	`"qwen3"`	`"llama"`	Required for correct model loading
`vocab_size`	`151936`	`128256`	LLaMA 3.1 vocabulary size
`bos_token_id`	`151643`	`128000`	LLaMA 3.1 BOS token
`eos_token_id`	`151645`	`128001`	LLaMA 3.1 EOS token
`num_hidden_layers`	`36`	`32`	LLaMA 3.1 8B has 32 layers
`intermediate_size`	`12288`	`14336`	Correct FFN dimension for LLaMA 3.1 8B
`rms_norm_eps`	`1e-06`	`1e-05`	LLaMA standard value
`rope_scaling`	`null`	`{...llama3 config}`	LLaMA 3.1 uses RoPE scaling
`rope_theta`	`1000000`	`500000.0`	LLaMA 3.1 8B value

generation_config.json

Field	Original	Fixed	Reason
`bos_token_id`	`151643`	`128000`	Match LLaMA tokenizer
`eos_token_id`	`[151645, 151643]`	`[128001, 128008, 128009]`	LLaMA 3.1 EOS tokens
`pad_token_id`	`151643`	`128004`	LLaMA finetune pad token

New Files Added

tokenizer_config.json - Required for chat template and proper tokenization
special_tokens_map.json - Defines BOS, EOS, PAD token strings
test_broken_model.py - Validation script to verify fixes

Validation

Run the test script to verify the configuration:

bash

pip install transformers torch
python test_broken_model.py

Test Output

============================================================
Broken Model Configuration Test Suite
============================================================

Testing model at: /Users/satish/Desktop/tasks/broken-model

=== Testing config.json ===
  ✓ architectures: ['LlamaForCausalLM']
  ✓ model_type: llama
  ✓ vocab_size: 128256
  ✓ bos_token_id: 128000
  ✓ eos_token_id: 128001
  ✓ num_hidden_layers: 32
  ✓ hidden_size: 4096
  ✓ num_attention_heads: 32
  ✓ num_key_value_heads: 8

=== Testing tokenizer_config.json ===
  ✓ bos_token: <|begin_of_text|>
  ✓ eos_token: <|eot_id|>
  ✓ chat_template: present (4648 chars)
  ✓ tokenizer_class: PreTrainedTokenizerFast

=== Testing special_tokens_map.json ===
  ✓ bos_token: <|begin_of_text|>
  ✓ eos_token: <|eot_id|>

=== Testing generation_config.json ===
  ✓ bos_token_id: 128000
  ✓ eos_token_id: [128001, 128008, 128009]

=== Testing transformers loading ===
  ✓ AutoConfig loaded successfully
    - Architecture: ['LlamaForCausalLM']
    - Model type: llama
    - Vocab size: 128256
  ✓ Model correctly identified as LLaMA

=== Testing model architecture initialization ===
  ✓ Config validated for LlamaForCausalLM
    - Layers: 32
    - Hidden size: 4096
    - Vocab size: 128256
    - Estimated params: ~8.8B

============================================================
Test Summary
============================================================
  ✓ PASS: config.json
  ✓ PASS: tokenizer_config.json
  ✓ PASS: special_tokens_map.json
  ✓ PASS: generation_config.json
  ✓ PASS: transformers loading
  ✓ PASS: model initialization

Result: 6/6 tests passed

✓ All tests passed! The model configuration is valid.

Usage

These config files can be used with actual LLaMA 3.1 8B weights:

python

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load with official weights
model = AutoModelForCausalLM.from_pretrained("meta-llama/Meta-Llama-3.1-8B")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3.1-8B")

Or use these corrected config files with your own fine-tuned weights.

Files

File	Description
`config.json`	Model architecture configuration
`generation_config.json`	Generation parameters
`tokenizer_config.json`	Tokenizer settings with chat template
`special_tokens_map.json`	Special token definitions
`merges.txt`	BPE merge rules
`test_broken_model.py`	Validation test script

devsatish/fixed_model