Modelpublic

marcodsn/qwen3.5-4b-sft-flint-section-v2-3ep

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

0likes

Model Card

qwen3.5-4b-sft-flint-section-v2-3ep

sft-flint-section-v2-3ep arm of the caveman reasoning-compression ablation study: Qwen/Qwen3.5-4B fine-tuned (LoRA adapter) on flint/data/flint-section-aware.jsonl (552 rows, 3 epochs, LoRA r=64).

The study asks whether compressed ("caveman") reasoning traces can train a model to reason in fewer tokens without losing accuracy — and which parts of a trace are compressible. See the run manifest below for the exact recipe; eval results live in the study's report.

Eval summary (t=0, max_tokens 8192)

Accuracy (avg reasoning tokens, loop rate) — this arm vs the original model it was fine-tuned from (Qwen/Qwen3.5-4B), same harness and prompts.

suite	this model	original Qwen3.5-4B
gsm8k@t0.0	0.66 (3483.3 tok, loops 0.22)	0.575 (4413.4 tok, loops 0.2167)
math500@t0.0	0.49 (5472.3 tok, loops 0.18)	0.3083 (6631.6 tok, loops 0.2833)

Run manifest

json

{
  "arm": "sft-flint-section-v2-3ep",
  "dataset": "flint/data/flint-section-aware.jsonl",
  "rows": 552,
  "dropped_overlong": 0,
  "epochs": 3,
  "system_prompts": true,
  "lora": {
    "r": 64,
    "alpha": 128,
    "dropout": 0.0,
    "target": "all"
  },
  "train": {
    "epochs": 2,
    "lr": 0.0002,
    "batch_size": 1,
    "grad_accum": 16,
    "warmup_ratio": 0.03,
    "lr_scheduler": "cosine",
    "weight_decay": 0.01,
    "seed": 3407,
    "logging_steps": 10,
    "save_strategy": "epoch"
  },
  "model": {
    "name": "Qwen/Qwen3.5-4B",
    "max_seq_length": 13312,
    "load_in_4bit": true,
    "chat_template": "qwen3.5"
  },
  "train_runtime_s": 7129.7994,
  "final_loss": 0.18160040378570558,
  "log_history": [
    {
      "loss": 0.3623380422592163,
      "grad_norm": 0.20898976922035217,
      "learning_rate": 0.0001987930439740757,
      "epoch": 0.2898550724637681,
      "step": 10
    },
    {
      "loss": 0.3434292793273926,
      "grad_norm": 0.19094254076480865,
      "learning_rate": 0.00018931150161867916,
      "epoch": 0.5797101449275363,
      "step": 20
    },
    {
      "loss": 0.34574434757232664,
      "grad_norm": 0.16883905231952667,
      "learning_rate": 0.00017125839641475072,
      "epoch": 0.8695652173913043,
      "step": 30
    },
    {
      "loss": 0.26570911407470704,
      "grad_norm": 0.15527313947677612,
      "learning_rate": 0.00014636635319853275,
      "epoch": 1.144927536231884,
      "step": 40
    },
    {
      "loss": 0.23108193874359131,
      "grad_norm": 0.16759531199932098,
      "learning_rate": 0.00011702435557223987,
      "epoch": 1.434782608695652,
      "step": 50
    },
    {
      "loss": 0.23632392883300782,
      "grad_norm": 0.16645948588848114,
      "learning_rate": 8.604846610560771e-05,
      "epoch": 1.7246376811594204,
      "step": 60
    },
    {
      "loss": 0.24004652500152587,
      "grad_norm": 0.2916702926158905,
      "learning_rate": 5.6411558152462894e-05,
      "epoch": 2.0,
      "step": 70
    },
    {
      "loss": 0.17817842960357666,
      "grad_norm": 0.16276057064533234,
      "learning_rate": 3.0957997942825336e-05,
      "epoch": 2.289855072463768,
      "step": 80
    },
    {
      "loss": 0.17362045049667357,
      "grad_norm": 0.18809917569160461,
      "learning_rate": 1.2130659990073146e-05,
      "epoch": 2.579710144927536,
      "step": 90
    },
    {
      "loss": 0.18160040378570558,
      "grad_norm": 0.16632623970508575,
      "learning_rate": 1.7364751777736332e-06,
      "epoch": 2.869565217391304,
      "step": 100
    },
    {
      "train_runtime": 7129.7994,
      "train_samples_per_second": 0.232,
      "train_steps_per_second": 0.015,
      "total_flos": 2.0013017119053005e+17,
      "train_loss": 0.2515000207083566,
      "epoch": 3.0,
      "step": 105
    }
  ]
}