SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2407.10490 · 2024

Learning Dynamics of LLM Finetuning

arXiv · PDF · Open in the Atlas

Code that ran

7 of 10 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 7 failures of our instrument; 3 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2407.10490.json · badge: /api/ran/2407.10490.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
joshua-ren/learning_dynamics_llm canonical (classifier label)
classifier basis: census
7 ran · 3 no recorded run
FunctionStatusWhere it lives
extract_anthropic_prompt Ran
ran; our contract draft was wrong, not the code
joshua-ren/learning_dynamics_llm/src/preference_datasets.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from junkangwu/beta-DPO, is pointer only (licence: NONE) · get_code("12f6803880f9b17e")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from junkangwu/beta-DPO, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

get_api_base_url Ran
ran; our contract draft was wrong, not the code
joshua-ren/learning_dynamics_llm/src/gen_dataset.py
code served (permissive licence) · get_code("950dcf3ad144388a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
import os
from pathlib import Path

# Test case 1: Using default platform (openai) without api_base_url.txt file
ARGS = {
    'platform': 'openai'
}

# Alternative test case 2: Using baichuan platform without api_base_url.txt file
# ARGS = {
#     'platform': 'baichuan'
# }
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_hh Ran
ran; our contract draft was wrong, not the code
joshua-ren/learning_dynamics_llm/src/preference_datasets.py
code served (permissive licence) · get_code("7d82ad82e19dab04")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json
import tempfile
import os

# Create a temporary directory for test data
temp_dir = tempfile.mkdtemp()

# Create a sample JSONL file for the 'train' split
train_data = [
    {
        'prompt': 'H: What is 2+2?\n\nA:',
        'chosen': '2+2 equals 4.',
        'rejected': '2+2 equals 5.',
        'random': '2+2 equals 3.',
        'paraphrase': 'The sum of two and two is four.',
        'variant': '2 plus 2 is 4.',
        'nonresponse': 'I cannot answer this.'
    },
    {
        'prompt': 'H: How are you?\n\nA:',
        'chosen': 'I am doing well, thank you for asking.',
        'rejected': 'I am doing poorly.',
        'random': 'I am neutral.',
        'paraphrase': 'I am in good spirits.',
        'variant': 'Things are going well for me.',
        'nonresponse': 'No response available.'
    }
]

# Write the test data to a JSONL file
with open(f'{temp_dir}/train.jsonl', 'w') as f:
    for item in train_data:
        f.write(json.dumps(item) + '\n')

# Define the arguments for the get_hh function
ARGS = {
    'split': 'train',
    'silent': False,
    'data_dir': temp_dir
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
tqdm — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_local_dir Ran
ran; our contract draft was wrong, not the code
joshua-ren/learning_dynamics_llm/src/utils.py
code served (permissive licence) · get_code("8ab43db002bf984d")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import os
import tempfile
import getpass

# Create a temporary directory that exists for testing
temp_dir = tempfile.mkdtemp()

# Create the ARGS dictionary with valid arguments for get_local_dir
ARGS = {
    'prefixes_to_resolve': [temp_dir, '/tmp', '/var/tmp']
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_local_run_dir Ran
ran; our contract draft was wrong, not the code
joshua-ren/learning_dynamics_llm/src/utils.py
code served (permissive licence) · get_code("10ab71f27ffd7de8")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import os
import tempfile
from datetime import datetime

# Create a temporary directory for testing
temp_base = tempfile.mkdtemp()

ARGS = {
    'exp_name': 'test_experiment',
    'local_dirs': [temp_base, '/nonexistent/path']
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_remote_file Ran
ran; our contract draft was wrong, not the code
joshua-ren/learning_dynamics_llm/src/utils.py
code served (permissive licence) · get_code("e7ce4c19888c6df6")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import os
import socket
import tempfile

# Create a temporary directory for testing
temp_dir = tempfile.mkdtemp()

# Get the current hostname
current_hostname = socket.gethostname()

# Create a test local path
test_local_path = os.path.join(temp_dir, "test_file.txt")

# Create a remote path string - using local hostname to avoid actual SCP calls
# Format is "hostname:path"
remote_path = f"{current_hostname}:/tmp/test_remote_file.txt"

ARGS = {
    "remote_path": remote_path,
    "local_path": test_local_path,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
preference_loss Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
joshua-ren/learning_dynamics_llm/src/trainers.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from icip-cas/Word-Deed-Consistency-Test, is pointer only (licence: NONE) · get_code("c79211fc0225a27f")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from icip-cas/Word-Deed-Consistency-Test, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

get_dataset Not yet run
unverified (harvested, no recorded run)
joshua-ren/learning_dynamics_llm/src/preference_datasets.py
code served (permissive licence) · get_code("8fa0d07c713b5722")
How it ran

no invocation recorded.

get_gpt_completion Not yet run
unverified (harvested, no recorded run)
joshua-ren/learning_dynamics_llm/src/gen_dataset.py
code served (permissive licence) · get_code("d6be0aae15f087e3")
How it ran

no invocation recorded.

get_original_helpful Not yet run
unverified (harvested, no recorded run)
joshua-ren/learning_dynamics_llm/src/gen_dataset.py
code served (permissive licence) · get_code("a5e707e0c89d915f")
How it ran

no invocation recorded.

Also present in other papers' code

Of 10 samples lifted from this paper's linked repositories, 10 carry a function identity and 0 do not. Of the 10 distinct functions with an identity, 5 are also present in other papers whose code we hold (5 with at least one paper dated earlier, 5 the same month, 5 later) and 5 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

FunctionAlso present inEarliest-datedOthers named
extract_anthropic_prompt 15 other papers
11 dated earlier, 2 the same month, 2 later
2309.16240 (2023-09)
Beyond Reverse KL: Generalizing Direct Preference Optimization with Diverse Divergence Constraints
2310.08164 (2023-10)
Interpreting Learned Feedback Patterns in Large Language Models
2311.08380 (2023-11)
Direct Preference Optimization for Neural Machine Translation with Minimum Bayes Risk Decoding
and 12 more not named here
get_remote_file 15 other papers
10 dated earlier, 2 the same month, 3 later
2309.16240 (2023-09)
Beyond Reverse KL: Generalizing Direct Preference Optimization with Diverse Divergence Constraints
2312.00267 (2023-12)
Sample Efficient Preference Alignment in LLMs via Active Exploration
2401.01967 (2024-01)
A Mechanistic Understanding of Alignment Algorithms: A Case Study on DPO and Toxicity
and 12 more not named here
get_local_run_dir 11 other papers
8 dated earlier, 2 the same month, 1 later
2309.16240 (2023-09)
Beyond Reverse KL: Generalizing Direct Preference Optimization with Diverse Divergence Constraints
2312.00267 (2023-12)
Sample Efficient Preference Alignment in LLMs via Active Exploration
2401.13275 (2024-01)
Can AI Assistants Know What They Don't Know?
and 8 more not named here
get_local_dir 10 other papers
6 dated earlier, 2 the same month, 2 later
2309.16240 (2023-09)
Beyond Reverse KL: Generalizing Direct Preference Optimization with Diverse Divergence Constraints
2401.01967 (2024-01)
A Mechanistic Understanding of Alignment Algorithms: A Case Study on DPO and Toxicity
2401.13275 (2024-01)
Can AI Assistants Know What They Don't Know?
and 7 more not named here
preference_loss 5 other papers
3 dated earlier, 1 the same month, 1 later
2305.18290 (2023-05)
Direct Preference Optimization: Your Language Model is Secretly a Reward Model
2401.13275 (2024-01)
Can AI Assistants Know What They Don't Know?
2405.20304 (2024-05)
Group Robust Preference Optimization in Reward-free RLHF
and 2 more not named here

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Learning dynamics, which describes how the learning of specific training examples influences the model's predictions on other examples, gives us a powerful tool for understanding the behavior of deep learning systems. We study the learning dynamics of large language models during different types of finetuning, by analyzing the step-wise decomposition of how influence accumulates among different potential responses. Our framework allows a uniform interpretation of many interesting observations about the training of popular algorithms for both instruction tuning and preference tuning. In particular, we propose a hypothetical explanation of why specific types of hallucination are strengthened after finetuning, e.g., the model might use phrases or facts in the response for question B to answer question A, or the model might keep repeating similar simple phrases when generating responses. We also extend our framework and highlight a unique "squeezing effect" to explain a previously observed phenomenon in off-policy direct preference optimization (DPO), where running DPO for too long makes even the desired outputs less likely. This framework also provides insights into where the benefits of on-policy DPO and other variants come from. The analysis not only provides a novel perspective of understanding LLM's finetuning but also inspires a simple, effective method to improve alignment performance.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2407.10490")
get_code_for_paper("2407.10490")
have("2407.10490")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2407.10490.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2407.10490.svg)](https://syntology.ai/paper/2407.10490)

Connect an agent — have() is free.