Qwen2.5-7B-Instruct mixed-v9 SFT
This is the selected epoch-1 (checkpoint-66) full-SFT checkpoint used by the
mixed-v9 profile → interpretation → response experiments.
The model was initialized from Qwen/Qwen2.5-7B-Instruct and trained with
LLaMA-Factory on 2,092 mixed-v9 rows:
- 1,192 profile examples: user history → structured user profile;
- 450 interpretation rows (150 distinct examples, upsampled 3×): profile plus query/reply/feedback → five-field interpretation;
- 450 response rows (150 distinct examples, upsampled 3×): profile plus interpretation → corrected response.
Checkpoint 66 was selected using the disjoint mixed-v9 development set because its development loss was lower than checkpoint 132 (1.1218 vs. 1.2826).
Matched held-out evaluation
On the 100-user mixed-v9 evaluation set, using GPT-4o as the fixed judge, this checkpoint improved over the original base checkpoint as follows:
| Metric | Base | This checkpoint | Paired difference |
|---|---|---|---|
| Profile | 0.3758 | 0.6233 | +0.2475 |
| Interpretation | 0.2430 | 0.3850 | +0.1420 |
| Response | 0.1745 | 0.3272 | +0.1527 |
| RL-weighted total | 0.1026 | 0.1772 | +0.0746 |
These results measure the complete staged task and should not be interpreted as the causal value of the user profile alone. The companion ablation repository separates direct response generation, profile-only, interpretation-only, and full staged generation.
Loading
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "Aaron757/Qwen2.5-7B-Instruct-v9-mixed-SFT"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype="auto",
device_map="auto",
)
For vLLM, point --model at this repository ID. The native context length is
32,768 tokens.
Scope and limitations
- The checkpoint is specialized for the exact structured prompts used by the three-stage pipeline; direct-response prompts are an out-of-distribution diagnostic unless a matched direct-response model is trained.
- The interpretation references are model-generated pseudo-labels accepted by the v7 verification pipeline, not human-verified intent labels.
- The training data is primarily English WildChat-derived feedback data. Do not assume the profile or interpretation is factually correct without checking the supplied evidence.
- The upload contains inference weights and tokenizer/configuration files only; DeepSpeed optimizer states are intentionally excluded.